The Georgia Tech Class of 2026 celebrated one of its most significant accomplishments last week. Among those crossing the stage for the 2026 Spring Commencement is Marina Vemmou, who will be receiving her Ph.D. in computer science.   

As a machine learning (ML) theorist, Xinyuan Cao spent her Ph.D. at Georgia Tech digging into the mathematical foundations of artificial intelligence (AI). 

TokenSmith is a citation-supported large language model (LLM) tutor that can be hosted locally on a user’s personal computer. The tutor only provides answers based on course materials, such as the textbook or lecture slides.  

Associate Professor Joy Arulraj began the project with support from the Bill Kent Family Foundation AI in Higher Education Faculty Fellowship last year. The fellowship, led by Georgia Tech’s Center for 21st Century Universities, supports faculty projects exploring innovative and ethical uses of AI in teaching.   

Arulraj has enlisted assistant professors Kexin Rong and Steve Mussmann to help build TokenSmith.  

Mussmann said TokenSmith is a synergistic blend of a database system and a machine learning system. The model stores textbooks, textbook annotations by course staff, common questions and answers, a learning state of the student, and student feedback in a structured database system. However, machine learning plays a key role in the answer generation as well as adapting the system to the student, course staff guidance, and user feedback.

"What excites me most is demonstrating how data-driven ML and principled database systems design can reinforce each other — one providing adaptability and flexibility, the other providing structure and traceability — in a way that benefits students," Mussmann said.

Keeping the model local has been an important focus of the project. The team wanted to create an AI tutor that helps students learn from their class resources rather than just giving answers. With each response, TokenSmith cites the origin of the answer in the provided documents.  

“One problem with LLMs is that they can hallucinate and provide wrong answers, but in this controlled environment, we can add these guardrails to make sure it’s actually helpful in an educational setting,” Rong said.  

Rong said she feels that students often undervalue textbooks, and she hopes TokenSmith can motivate students to make better use of them.  

“Textbooks can sometimes be daunting, but maybe if we combine them with the model, students might be more willing to read a paragraph or page in the textbook, and that could help clarify something for them,” she said.  

Running the model locally is more cost-effective and helps preserve the user’s privacy. But running the new tool locally comes with technical challenges.  

One challenge with creating the model is speed. Since it is a locally based model, TokenSmith depends solely on the user’s computer memory.  Tests have also shown that the tutor currently struggles to answer more complex questions. 

“We are interested in pushing the boundaries of these local models so that they give students good answers and also run fast enough to keep students engaged,” Arulraj said.  

When the first cryptocurrency, Bitcoin, was proposed in 2008, the goal was simple: to create a digital currency free from banks and governments. Over time, that idea evolved into something much bigger: “decentralized finance,” or “DeFi.”

With decentralized finance, people trade, borrow and earn interest on crypto assets without relying on traditional intermediaries. DeFi services run on blockchains, which are essentially digital ledgers, and use “smart contracts” − self-executing code that automates financial transactions. Tens of billions of dollars have poured into the DeFi market.

But with innovation comes risks. The lack of centralized oversight has made crypto, including decentralized finance, a prime target for hackers and scammers. In 2024 alone, people lost nearly US$1.5 billion due to security exploits and fraud. And unlike traditional finance, there’s usually no way to recover stolen crypto.

As a computer scientist, I wanted to better understand how people perceive and respond to these risks. So my colleagues and I first conducted in-depth interviews with 14 crypto investors, then surveyed nearly 500 others to validate our findings.

Our study found that people often made the same mistakes, driven by recurring misconceptions and gaps in security awareness. Here are some of the most important.

Mistake 1: Thinking the blockchain guarantees security

Many people told us they thought decentralized finance was secure – but their reasoning wasn’t very convincing. Some seemed to confuse decentralized finance with blockchain technology itself, which is designed to ensure transactions are tamper-resistant through so-called “consensus mechanisms.” One told us that DeFi is secure “because a hacker would have to override an entire blockchain” to steal funds.

But services on the blockchain are still vulnerable to implementation and design flaws. These include smart contract breaches, in which bad guys exploit bugs in a service’s code, and front-end attacks, where a user interface is altered to redirect funds into a hacker’s wallet. A front-end attack was reportedly to blame for a recent $1.5 billion crypto heist.

Mistake 2: Thinking safe keys mean safe funds

Another common misconception is that DeFi is secure if private keys are well stored. A private key is a secret code that allows someone to access their crypto assets. It’s true that in DeFi – unlike in centralized crypto finance where an exchange holds private keys – users have full control over their own private keys.

But even with perfect private key management, users can still lose funds by interacting with compromised DeFi platforms. That’s because safeguarding private keys can prevent only direct attacks targeting private key access, such as phishing attempts.

The people we spoke with also failed to follow best practices for securing their private keys. Using a hardware wallet – a physical device that stores private keys offline – is one of the most secure options for protecting keys from online threats. However, our study found that only a handful of participants actually used hardware wallets.

Mistake 3: Thinking 2-factor authentication is a silver bullet

Two-factor authentication, or 2FA, is a standard security mechanism in which two forms of verification are required to access an account. Think being texted a one-time code before you can log into your bank account.

To prevent account breaches, centralized crypto exchanges such as Binance and Coinbase use two-factor authentication for logins, account recovery and withdrawal confirmations. But while 2FA is crucial to security in the traditional and centralized crypto finance system, it plays a much smaller role in decentralized finance.

DeFi wallets give users access based on private key ownership rather than identity verification, which means traditional 2FA can’t be used. Instead, only 2FA-like mechanisms are available in DeFi. For instance, multisignature wallets require approval from multiple private key holders. However, if your private key is compromised, attackers can perform wallet operations on your behalf without any additional verification. In addition, even users who adopt 2FA-like measures can’t prevent the security breaches on the DeFi services’ end.

Unfortunately, our participants were overly confident regarding the effectiveness of 2FA, with one saying, “Two-factor authentication has been one of the best solutions for keeping wallets safe.” In our survey, 57.1% of users relied on 2FA as their only technical countermeasure against rug pulls – scams where project creators suddenly withdraw funds – and 49.3% did so for smart contract exploits. This misplaced trust could lead them to ignore more effective security strategies.

Mistake 4: Not managing token approvals

One such effective strategy is revoking token approvals. In DeFi, tokens are digital assets on a blockchain that represent value or rights, and users often need to approve smart contracts to access or spend them. But if you leave these approvals open, a malicious contract – or one that’s been hacked – can drain your wallet. So it’s crucial to routinely check all token approvals you’ve granted to prevent losses caused by fraudulent or hacked DeFi services. Specifically, you should limit spending allowances instead of using the default “unlimited” option, and revoke approvals for apps you no longer use or trust.

Worryingly, we found that only 10.8% and 16.3% of participants regularly checked and revoked token approvals to protect against rug pulls and smart contract exploits, respectively. In light of this, we recommend that wallet providers introduce a reminder feature to prompt users to review their token approvals periodically.

Mistake 5: Not learning from past incidents

Even after they’re hacked or scammed, people often don’t do anything to improve their security practices, we found. Just 17.6% of those who reported being victims of a DeFi scam regularly checked token approvals afterward. Worse, 26% took no action at all after a scam, and 16.4% doubled down by investing even more in other DeFi services.

Surprisingly, more than half of the victims said their belief in DeFi either stayed the same or grew stronger after the incident. One user who lost $4,700 due to a rug-pull incident said, “My belief in cryptocurrency has grown stronger after that because I made good money from it.” That person added, “An opportunity to make money is something I believe in.” This suggests that DeFi users’ financial motivations can sometimes outweigh their security concerns – and, perhaps, their better judgment.

There’s no one-size-fits-all solution to DeFi security. But awareness is the first step. To stay safe, crypto investors should use hardware wallets, revoke unused token approvals and continually learn new techniques to protect themselves from evolving threats. Most importantly, they should stay rational and not let the allure of profits cloud their security practices.

This article is republished from The Conversation under a Creative Commons license. Read the original article.

Socrates, widely regarded as the founder of Western philosophy, declared, "I know that I know nothing." This simple statement highlights the wisdom of acknowledging the limits of one's own knowledge.

A simple statement, yes, but like some people, LLMs struggle with saying “I don’t know.” In fact, LLMs often can't admit that they don't know something because of the way they are trained, according to a research team that includes a Georgia Tech computer science (CS) professor.

Pre-training LLMs involves them learning to predict the next word correctly by training on massive datasets of text, images, or other data. Models are evaluated and adjusted based on their performance against standard benchmarks, which are "rewarded" for preferred outputs or answers.

Current evaluation protocols, however, penalize non-responses the same as incorrect answers and do not include an "I don't know" option.

According to CS Professor Santosh Vempala, these pre- and post-training shortcomings are what lead LLMs to provide seemingly plausible but false responses known as hallucinations.

Vempala is a co-author of Why Language Models Hallucinate, a research study from OpenAI and Georgia Tech, released in September. He says that there is a direct correlation between an LLM's hallucination rate and its misclassification rate regarding the validity of a given response.

"This means that if the model can't tell fact from fiction, it will hallucinate," Vempala said. 

"The problem persists in modern post-training methods for alignment, which are based on evaluation benchmarks that penalize 'I don't know' as much as wrong answers."

Because of the penalties for knowing that it knows nothing – to paraphrase Socrates – guessing is a more rewarding option for current LLMs than admitting uncertainty or ignorance.

The research incorporates and builds on prior work from Vempala and Adam Kalai, an OpenAI researcher and lead author of the current paper. Their earlier work found that LLM hallucinations are mathematically unavoidable for arbitrary facts, given current training methodologies. 

"We've been talking about this for about two years. One corollary of our paper is that, for arbitrary facts, despite being trained only on valid data, the hallucination rate is determined by the fraction of missing facts in the training data," said Vempala, Frederick Storey II Chair of Computing and professor in the School of CS.

To illustrate this point, imagine you have a huge Pokémon card collection. Pikachu is so familiar that you can confidently describe its moves and abilities. However, accurately remembering facts about Pikachu Libre, an extremely rare card, would likely be more difficult.

“More to the point, if your collection has a large fraction of unique cards, then it is likely that you are still missing a large fraction of the overall set of cards. This is known as the Good-Turing estimate,” Vempala said.

[OpenAI Blog: Why Language Models Hallucinate]

According to Kalai and Vempala, the same is true for LLMs based on current training protocols.

“Think about country capitals,” Kalai said. “They all appear many times in the training data, so language models don’t tend to hallucinate on those.

“On the other hand, think about the birthdays of people’s pets. When those are mentioned in the training data, it may just be once.

“So, pre-trained language models will hallucinate on those. However, post-training can and should teach the model not to guess randomly on facts like those.”

Vempala thinks tinkering with pre-training methods could be risky because, overall, they work well and deliver accurate results. However, he and his co-authors offered suggestions for reducing the occurrence of hallucinations with changes to the evaluation and post-training process.

Among the team's recommended changes is that more value be placed on the accuracy of an LLM's responses rather than on how comprehensive their responses are. The team also suggests implementing what it refers to as “behavioral calibration.”

Using this methodology, LLMs would only answer if their confidence level exceeds target thresholds. These thresholds would be tuned for different user domains and prompts. They would also appropriately reduce penalties for “I don’t know” responses, along with appropriate expressions of uncertainty and wrong answers.

Vempala believes that implementing some of these modifications could result in LLMs that are trained to be more cautious and truthful. This shift could lead to more intelligent systems in the future that can handle nuanced, real-world conversations more effectively.

"We hope our recommendations will lead to more trustworthy AI," said Vempala. "However, implementing these modifications to how LLMs are currently evaluated will require acceptance and support from AI companies and users."

One of these researchers is Amanda Meng, senior research scientist in the School of Computer Science (SCS). The overarching theme in Meng’s work is the relationship between power and data and how different social groups can make use of data to shift power. 

As the only social scientist in SCS, Meng sees her role as an “important and potentially powerful interdisciplinary connection.” 

Connecting Social Justice with Data 

Although focused on political and social change, Meng’s work has always had links to technology. 

After completing her undergraduate education at Georgia Tech, Meng joined the Peace Corps, where she served in the Dominican Republic. She spent two years there working to improve computer literacy in schools and create community computer labs. 

Meng said her time in the Peace Corps made her interested in how communities advocated for themselves. She explored this idea further while completing her Ph.D. from the Georgia Tech Sam Nunn School of International Affairs. 

With her Ph.D. in hand, Meng was hired as a research scientist in SCS, working under Professor Ellen Zegura and School of Interactive Computing Professor Carl DiSalvo on civic data projects based in Atlanta. 

This experience made her curious about the interaction between data literacy and civic literacy. 

“We live in such a data-fied society that a lot of advocacy work often does involve data because to make your claims legitimate, policy makers want to see and understand the data,” she said. 

Following a brief stint in the private sector as a data consultant, Meng returned to SCS, this time as a research scientist working on IODA (Internet Outage Detection and Analysis) with Associate Professor Alberto Dainotti. IODA is a research project and online platform that provides real-time measurements on global internet connectivity. 

In her contribution to the IODA project, Meng aims to improve the usability of IODA, particularly by users affected by government-ordered shutdowns, by developing IODA users’ internet measurement literacy. Currently, IODA provides the most granular, near-real-time data on Internet infrastructure connectivity. Meng uses this data to collaborate with global advocacy groups to publish reports detailing IODA’s measurements alongside its sociopolitical context. Meng said the eventual goal of her work with IODA is for others to know how to use the platform to monitor for events and advocate against shutdowns. 

“The platform is really only as successful as its userbase is at understanding, making use, and acting on its data,” Meng said. 

In the past year, Meng was awarded her first grant as principal investigator. The grant uses Aggie, an open-source tool developed at Tech that aggregates content from the internet. 

Previously, Aggie has been used to monitor elections on social media. Meng said she wants to explore using it to monitor internet shutdowns or censorship events. She is currently conducting a pilot study to test the system, which will determine whether Aggie offers a more collaborative and coordinated way to monitor connectivity across measurement and social media data. 

The Role of a Research Scientist 

As a research scientist, almost all of Meng’s work is oriented toward research. This includes working on proposals and existing grants, as well as advising students. 

“We love to involve students in our research,” she said. “We aren’t just here to do research. We are here to involve students in research so they can learn and develop domain skills and research skills.” 

Since 2024, Meng has served on the School Advisory Committee in SCS. She says it’s important to have research faculty in service roles, as they have a different set of needs in their position. 

“Through the funding we can apply for, the research we do, and the work we do with students, we are an important multiplier for the work that the School wants to cultivate,” Meng said. 

Community Driven 

Even as a Ph.D. student, Meng said she wanted to contribute to local community groups. An important value she learned in the Peace Corps was to be an active participant in the community she lived in. 

Meng started getting involved in a housing justice project in Atlanta’s Westside neighborhood that was collecting data related to their mission. They soon discovered that some residents’ data was more accurate than official records because they lived there. 

“We also learned it’s not all about impacting legislation,” Meng said. “It’s about mobilizing resources within the community, and the fact that data could be used to do that was an important finding, and it’s something that I want to continue to draw out with data and AI.” 

Meng has continued to work with the group from that project and wants to continue ethnographic research into how data and AI are used to create change. 

“AI could have the ability to consolidate power in the hands of those who develop closed-source models," Meng said. "It’s important to study the entities that are developing AI as much as we study the communities that might make use of or be most minoritized by AI." 

The innovative model, developed at Georgia Tech, creates invisible digital masks for personal photos to thwart unwanted online facial recognition while preserving the image quality.

Anyone who posts photos of themselves risks having their privacy violated by unauthorized facial image collection. Online criminals and other bad actors collect facial images by web scraping to create databases.

These illicit databases enable criminals to commit identity fraud, stalking, and other crimes. The practice also opens victims to unwanted targeted ads and attacks.

The new model is called Chameleon. Unlike current models, which produce different masks for each user’s photos, Chameleon creates a single, personalized privacy protection (P-3) mask for all of a user’s facial photos.

A bespoke P-3 mask is created based on a few user-submitted facial photos. After applying the mask, protected photos won’t be detectable by someone scanning for the user’s face. Instead, the unwanted scan will identify the protected photos as being someone else.

The Chameleon model was developed by Professor Ling Liu of the School of Computer Science (SCS), Ph.D. students Sihao Hu and Tiansheng Huang, and Ka-Ho Chow, an assistant professor at the University of Hong Kong and Liu’s former Ph.D. student.

During development, the team accomplished its two main goals: protecting the person's identity in the photo and ensuring a minimal visual difference between the original and masked photos.

The researchers said a notable visual difference often exists between the original and photos using current masking models. However, Chameleon preserves much of the original photo’s quality among various facial images.

In several research tests, Chameleon outperformed three top facial recognition protection models in visual and protective metrics. The tests also showed that Chameleon offers more substantial privacy protection while being faster and more resource-efficient.

In the future, Huang said they would like to apply Chameleon’s methods to other uses.

“We would like to use these techniques to protect images from being used to train artificial intelligence generative models. We could protect the image information from being used without consent,” he said.

The research team aims to release Chameleon code publicly on GitHub to allow others to improve their work.

“Privacy-preserving data sharing and analytics like Chameleon will help to advance governance and responsible adoption of AI technology and stimulate responsible science and innovation,” said Liu.

The paper on Chameleon, Personalized Privacy Protection Mask Against Unauthorized Facial Recognition, was presented earlier this month at ECCV 2024.

School of Computer Science Associate Professor Alberto Dainotti, however, is revolutionizing network telescopes through a $1.2 million grant from the National Science Foundation.

Network telescopes use large sets of inactive IP addresses to observe unsolicited internet traffic, typically considered “pollution,” to reveal many internet phenomena. These observations can be used to detect denial-of-service attacks and find viruses or other malicious activity.

Network telescopes' ability to monitor this pollution also provides a way to track internet connectivity. Network telescopes are one of the tools used by IODA, a system tracking connectivity worldwide created by Dainotti’s lab.

The larger and more accurate the telescope, the more inactive IP addresses it has. Due to the increasing cost and decreasing availability of IP addresses, creating and maintaining large network telescopes has become difficult for universities. Institutions have sold many of the addresses they own or allocated them to devices using the internet.

Dainotti will use his NSF grant to help universities and other organizations again have powerful network telescopes.

“If we stop seeing pollution coming from a particular area, maybe there’s something wrong with connectivity there since that pollution is typically happening constantly,” Dainotti said.

While universities might not have large numbers of inactive IP addresses to dedicate solely to a network telescope, many addresses aren’t always in use. Until now, it has not been easy to track this activity. However, Dainotti has created a system to detect this automatically. Using this method, organizations can create what Dainotti calls a dynamic network telescope.

The dynamic network telescopes also solve another problem: some malicious actors have learned how to detect and block the sets of IP addresses used in network telescopes. Using the dynamic approach makes it harder for them to track which addresses are currently being used.

“The spirit of this proposal is to reenable organizations to have this precious research infrastructure in a different way, but with the same purpose,” Dainotti said.

Ganesh and his student, John Lu, have been working on their own solver, Z3-alpha, for several years. It recently won several categories at SMT-COMP 2024, a competition held to determine the best solvers from around the world.

SMT solvers are automated logical reasoning tools used widely to test and analyze programs. They are also used to identify potential security issues.

“SMT solvers are like a Swiss Army Knife for all kinds of tasks for software engineering and trustworthy AI. You need a tool that can understand and analyze formulas obtained from analysis of programs,” Ganesh said.

The Z3-alpha solver is derived from the z3 solver from Microsoft Research, but Ganesh and Lu implemented machine learning (ML) into their solver to automatically synthesize strategies, making it more efficient.

Ganesh said the solver illustrates one of his biggest research goals: effectively combining the fields of automated reasoning and ML.

“With this solver, we’re using ML to make automated reasoning more efficient. In another line of research, we are going in the reverse direction by using automated reasoning to analyze, test, and improve ML models,” he said.

Using ML with SMT solvers this way is a relatively new line of research that Ganesh has been working on. He said this is among the first known instances of the successful use of machine learning for SMT solver strategy synthesis.

Ganesh said they want to work to further improve the Z3-alpha solver and apply these ML techniques to other solvers.

To counter these tactics and raise awareness, researchers with Georgia Tech’s Internet Intelligence Lab will use $1.4 million in new grants to advance the state-of-the-art in global internet connectivity detection, tracking, and analysis.

The Open Technology Fund awarded Associate Professor Alberto Dainotti and Research Scientists Zachary Bischof and Amanda Meng $960,840 to improve research infrastructure and methods to study internet censorship worldwide.

The funding supports efforts to improve the coverage and accuracy of IODA (Internet Outage Detection and Analysis). Developed in the Internet Intelligence Lab, IODA tracks internet outages worldwide caused by natural disasters, configuration issues, or censorship. IODA collects this data and provides it to the public on its website.

Over the years, IODA has provided data to the United Nations, Amnesty International, and other international organizations. Meng said one of their goals is to make IODA’s measurement data more accessible and usable for activists, intergovernmental organizations, and others in the internet freedom community.

“We want to enhance our tool to further enable them to monitor and track internet connectivity so that they can use that as evidence in their advocacy efforts,” she said.

Meng and her colleagues will also use the grant to pilot two projects with DataWorks. Housed in Georgia Tech’s College of Computing, DataWorks hires and trains people from the local community and provides data services to non-profit organizations.

DataWorks employees will help the researchers to:

• Pilot a global outage tracking team.
• Clean and complete historic outage dataset which contributed to groundbreaking research published at SIGCOMM on political and technical signatures of internet shutdowns.

The second grant, totaling $499,442, was awarded to Meng by the U.S. State Department. This funding supports a project to develop systems and processes that can provide a more coordinated and collaborative way of tracking censorship events in Iran. This award is Meng’s first federal grant as lead principal investigator.

Collaborating to Track Censorship

Along with these projects, the IODA team collaborates with other censorship measurement groups. Recently, the IODA team worked with M-Lab, who secured funding from the Open Technology Fund, to coordinate a series of meetings to bring the internet measurement community together to share research and best practices and learn from each other.

The first of these meetings convened at Georgia Tech in June. Along with the IODA team, attendees included:

• Censored Planet, a University of Michigan research group
• The Open Observatory for Network Interference
• M-Lab 
• Cloudflare

Meng says the overall goal of the first meeting was to identify ways in which these groups could coordinate more closely and become more knowledgeable about one another’s work.

“We are a set of groups that work together to verify outages or shutdowns as soon as they happen by comparing our data and making sure that we’re seeing similar patterns,” Meng said.

“While we work together in that way, we realized there was a lot that we didn’t know about each other’s datasets, and we could learn from each other in terms of infrastructure or how we go about detecting outages and censorship events.”

Georgia Tech will host the next meeting in December.

Van den Brand’s research focuses on dynamic algorithms and optimization algorithms.

Optimization algorithms support decision-making functionality in airline scheduling, automotive manufacturing, financial services, and other business processes.

Dynamic algorithms solve dynamic problems such as traffic congestion. When a GPS navigation app finds a route affected by construction or traffic congestion, it uses dynamic algorithms to quickly find new routes for the user.

“My research comes up with new techniques for how computers can maintain solutions for a problem that changes over time,” he said.

His proposal focuses on dynamic algorithms and linear programs, a type of optimization algorithm, and the relation between the two.

Although studied extensively, researchers know little about dynamic linear programs from a theoretical perspective. Van den Brand’s proposal explores the synergy between dynamic and optimization algorithms to build a theory for dynamic optimization problems.

Through this, van den Brand hopes to develop more general algorithm techniques that researchers can apply to many problems.

“It’s foundational research constructing tools and techniques for other engineers or scientists to solve their problems more efficiently,” van den Brand said.

As part of his proposal, he is launching a summer school program. The program will be a platform to introduce Ph.D. students to dynamic and optimization algorithms. Van den Brand says he hopes to introduce a new generation of students to theoretical computer science.

“I'm excited about the award and the new projects that it will fund. I want to thank NSF for the award and my colleagues for helpful advice during the proposal process,” van den Brand said.

Eye tracking is an essential component of AR/VR systems, but current systems have some limitations. These include a large size due to bulkier lens-based cameras and the high communication cost between the camera and the backend system.

Georgia Tech School of Computer Science Associate Professor Yingyan (Celine) Lin, Ph.D. student Hoaran You, and postdoctoral student Yang (Katie) Zhao have developed a new eye-tracking system that works around these limitations by combining a recently developed lens-less camera, algorithm, and acceleration processor designs.

“The current VR headsets are too heavy, gaming can lag, and using the controller is cumbersome. Combined, this prevents users from having a truly immersive experience. We mitigate all these problems,” said You.

This new system, EyeCoD: An Accelerated Eye Tracking System via FlatCam-based Algorithm & Accelerator Co-Design, replaces the traditional camera lens with FlatCam, a lensless camera 5x – 10x thinner and lighter. Combined with FlatCam, the team’s system enables eye tracking to function at a reduced size, with improved efficiency, and without sacrificing the accuracy of the tracking algorithm. The system could also enhance user privacy by not including a lens-based camera.

Another feature of the EyeCoD system is that it only puts the portion of the screen that a user’s eyes focus on in high resolution. It does this by predicting where a user’s eyes may land, then instantaneously rendering these areas in high res. These computational savings, plus a dedicated accelerator, underpin EyeCoD’s ability to boost processing speeds and efficiency.   

The team received the Office of Technology Licensing’s Tech Ready Grant for its efforts earlier this year. Tech Ready Grants offer $25,000 to help faculty transition projects from the lab to the marketplace.  

The team hopes to use the funds to integrate the current demos into a compact eye-tracking system for use in commercial VR/AR headsets.

Along with winning the Tech Ready Grant, the team presented EyeCoD at the International Symposium on Computer Architecture (ISCA) 2022. IEEE Micro included the work in its Top Picks from the Computer Architecture Conferences for 2023. The annual publication highlights “significant research papers in computer architecture based on novelty and potential for long-term impact.”  

EyeCoD is a collaborative work. Collaborators include Rice University Professor Ashok Veeraraghavan, whose team provided the technical support and design of the FlatCam camera in EyeCoD; and Ziyun Li, of Meta, who provided technical inputs to ensure that the EyeCoD system aligns with industry AR/VR specifications.  

A recent $623,790 grant from the Alfred P. Sloan Foundation is funding the creation of the Open Source Program Office (OSPO). The office will provide expert advice and the latest information to students and faculty interested in open-source research and publishing. The grant will support the office in its first two years.

Georgia Tech’s OSPO will be a virtual community served by members of the College of Computing, the Georgia Tech Library, the Center for Scientific Software Engineering, and the Partnership for an Advanced Computing Environment (PACE).

School of Computer Science Senior Research Scientist Jeffrey Young and PACE Senior Research Scientist Fang (Cherry) Liu will lead the new office. PACE Research Software Engineer Ron Rahaman, Interim Dean of Computing Alex Orso, and Susan Parham, Digital Curation Librarian at the Georgia Tech Library will provide support.

“If a student says, ‘I did something really cool for a class. I'd like to open-source it and get more users,’ we will give them advice on the best open-source license to pick, how to promote it to a community that would be interested in it, or how to package it in a way so that people can find it and use it easily,” said Young.

Along with answering questions about creating open-source software, contributing to existing open-source content, and curating open-source resources and tools, OSPO will develop a standard for open-source practices at Tech.

“Everybody knows or claims to know what open source is. The reality is that open source means so many different things to different people,” said Young. “We'd like to provide some guidance on developing your open-source work, encouraging external contributions to your project, and being consistent with university best practices for publishing and sharing your work.”

Young said the OSPO will work with the Georgia Tech Library to create and maintain an archive of open-source programs and develop strategies for preserving open-source projects for future use.

As part of their grant proposal, Young and Liu will also create a summer internship program to get students involved in open-source software and match students with faculty to create open-source projects.

Once the office is fully established, Young said they hope to engage with other groups to host workshops at Georgia Tech and other colleges and universities in the Atlanta area.

With support from philanthropic organizations like the Alfred P. Sloan Foundation, alumni, parents, friends, and corporations, Georgia Tech is securing the resources that will help achieve the most ambitious goals in the Institute's history as part of Transforming Tomorrow: The Campaign for Georgia Tech.

Elon University’s Imagining the Internet Center and the Pew Research Center published As AI Spreads, Experts Predict the Best and Worst Changes in Digital Life By 2035, on June 21. The 232-page report includes insights from more than 300 experts on how they imagine digital trends might change by 2035.

Along with Pu, Associate Professor Richard Barke from the School of Public Policy, and Digital Media Associate Professor Christopher Le Dantec, also represent Georgia Tech in the report.

When asked how they feel about possible developments in digital technology in the next decade, 42% of experts said they were equally excited and concerned. Thirty-seven percent said they were more concerned than excited.

Pu said he expects digital trends to be beneficial in many ways. But one of the main roadblocks in progress has been the standard ML practice of training models that are insensitive to evolutionary changes in time, making them obsolete at some point. Pu, who is the John P. Imlay, Jr. Chair in Software in the School of Computer Science, says future technology will likely overcome this obstacle. But he also thinks that as the power of technology increases, so will its influence, for better or worse.

“The power of AI/ML tools, particularly if they become sustainable and remain amoral, will be greater for both good and evil. We have seen significant harm from misinformation on the Covid-19 pandemic, dubbed as an ‘infodemic’ by the World Health Organization.

“More generally, it is being implemented in political propaganda in every election and every war. It is easy to imagine the depth, breadth, and constant renewal of such propaganda and infodemic, as well as their impact, all growing with the capabilities of future AI/ML tools used by powerful companies and governments,” said Pu, who is co-director of Georgia Tech’s Center for Experimental Research in Computer Systems.

Pu said AI/ML tools will likely become more widespread in the coming years. This might be an exciting idea for content producers, but it could be a problem when online users consume such content without critically thinking.

“The great digital divide is not going to be between the haves and have-nots of digital toys and information,” said Pu. “With more than six billion smartphones in the world (estimated in 2022), an overwhelming majority of the population already has access to and participates in the digital world.

“The digital divide in 2035 will be between those who think critically and those who believe misinformation and propaganda. This is a big challenge for democracy, a system in which we thought more information would be unquestionably beneficial. In a brave new digital world, a majority can be swayed by the misuse of amoral technological tools.

Born with cerebral palsy, Quinn has never let his condition slow him down, and he sees parallels between himself and the Institute’s iconic mascot. 

“I like the Wreck because I am a bit of a Ramblin’ Wreck myself. I self-manage my own team of caretakers who keep me running throughout the day, and I am infamous for speeding around campus in my wheelchair,” he said.  

Quinn always believed he could give the 1930 Ford Model A a run for its money in a race and got the opportunity to prove it. He led the Wreck throughout the race, from the starting line to the finish line at the end of Cherry Street. The speed of Quinn’s chair surprised Ramblin’ Wreck driver Matthew Kistner, but the master’s candidate has always approached life with a full-speed-ahead mentality.  

From the time he taught himself to use a computer in kindergarten, Quinn developed a passion for assistive technology and accessibility engineering. At 7 years old, he began working with Georgia’s Assistive Technology Act Program, Tools for Life, and went on to deliver keynote speeches across the state, including one at the University of Georgia. But he knew he wanted to go to Georgia Tech.  

Quinn was appointed as the youth advocate on the Tools for Life board following its adoption as a public service unit within the College of Design, where he began exploring the integration of artificial intelligence into human-centered design to enhance accessibility — a concept that he has been thinking about for some time.  

“AI will be one of the most important innovations in modern history for people with disabilities. Growing up, the running joke in my house was that I was going to live independently one day with the help of robots and self-driving cars. From autonomous vehicles hopefully making transportation more accessible, to generative AI models assisting with typing speed and productivity, the potential applications of AI for accessibility are truly endless,” he said.  

As an undergraduate, Quinn earned scholarships and awards from multiple organizations for academic achievements and leadership in disability advocacy. As the founding president of the ABLE Alliance (GTABLE), he has worked with Disability Services to advocate for students requiring more advanced accommodations and to secure independent living services through partnerships with state agencies. The alliance also hosted disability-focused professional development events with companies such as Google, Bloomberg, and Georgia Power.  

Quinn and GTABLE’s advocacy led to establishing a part-time dean’s list to honor students who cannot carry a minimum of 12 credit hours per semester due to extenuating life circumstances. 

“I am proud of the legacy of accessibility and inclusion I am leaving on campus,” he said. “My goal coming into Georgia Tech was to help make sure the Institute’s motto of progress and service was actively being practiced with respect to disability accommodations on campus. Through my work in disability advocacy and mentoring other up-and-coming disabled student leaders on campus, I am proud to have done my small part in making Georgia Tech work better for everyone.”  

After Commencement, Quinn will be a software engineer at Bloomberg, where he completed multiple internships. He plans to continue his advocacy work and hopes to one day lead product accessibility and inclusive design programs in the software industry.

The ability to uniformly scale the width (number of neurons) and depth (number of neural layers) of a DNN model means that remote clients can equitably participate in distributed, real-time training regardless of their computing resources. Resulting benefits include improved accuracy, increased efficiency, and reduced computational costs.

Developed by Georgia Tech researchers, the ScaleFL framework advances federated learning, which is an ML approach inspired by the personal data scandals of the past decade.

Federated learning (FL), a term coined by Google in 2016, enables a DNN model to be trained across decentralized devices or servers. Because data aren’t centralized with this approach, threats to data privacy and security are minimized.

The FL process begins with sending the initial parameters of a global DNN model to smartphones, IoT devices, edge servers, or other participating devices. These edge clients train their local version of the model using their unique data. All local results are aggregated and used to update the global model.

The process is repeated until the new model is fully trained and meets its design specifications.

Federated learning works best when remote clients involved in training a new DNN model have comparable computational power and bandwidth. But training can bog down if some participating remote-client devices have limited or fluctuating computing resources.

“In most real-life applications computational resources tend to differ significantly across clients. This heterogeneity prevents clients with insufficient resources from participating in certain FL tasks that require large models,” said School of Computer Science (CS) Ph.D. student Fatih Ilhan.

“Federated learning should promote equitable AI practice by supporting a resource-adaptive learning framework that can scale to heterogeneous clients with limited capacity,” said Ilhan, who is advised by Professor Ling Liu.

Ilhan is the lead author of ScaleFL: Resource-Adaptive Federated Learning with Heterogeneous Clients, which he is presenting at the 2023 Conference on Computer Vision and Pattern Recognition. CVPR 23 is set for June 18-22 in Vancouver, Canada.

Creating a framework that can adaptively scale the global DNN model based on a remote client’s computing resources is no easy feat. Ilhan says the balance between a model’s basic and complex feature extraction capabilities can be easily thrown out of whack when manipulating the number of neurons or the number of neuron layers of a DNN model.

“Since a deeper model is more capable of extracting higher order, complex features while a wider model has access to a finer resolution of lower-order, basic features, performing model size reduction across one dimension causes unbalance in terms of the learning capabilities of the resulting model,” said Ilhan.

The team overcomes these challenges in part by incorporating early exit classifiers into ScaleFL.

These ML-based tools are designed to optimize accuracy and efficiency by introducing intermediate decision points in the classification process. This capability enables a model to complete an inference task as soon as it is confident in its prediction, without having to process the whole model.

“ScaleFL injects these classifiers to the global model at certain layers based on the model architecture and computational constraints at each complexity level. This enables forming low-cost local models by keeping the layers up to the corresponding exit,” said Ilhan.

“Two-dimensional scaling with splitting the model along depth and width dimensions yields uniformly scaled, efficient local models for resource-constrained clients. As a result, not only does the global model achieves better performance compared to baseline FL approaches and existing algorithms, but local models at different complexity levels also perform significantly better for clients that are resource-constrained at inference time.”

The exit classifiers that help balance a model’s basic and complex features also play into the second part of ScaleFL’s secret sauce, self-distillation.

Self-distillation is a form of knowledge distillation, which has been used to transfer knowledge from a ‘teacher’ model to a smaller ‘student’ model. ScaleFL applies this process within the same network by comparing early predictions made by the exit classifiers (students) and the final predictions of the last exit (teacher) of local models during optimization. This technique prevents isolation and improves the knowledge transfer among subnetworks of different levels in ScaleFL.

Ilhan and his collaborators extensively tested ScaleFL on three image classification datasets and two natural language processing datasets.

“Our experiments show that ScaleFL outperforms existing representative heterogeneous federated learning approaches. In local model evaluations, we were able to reduce latency by two times, and the model size by four times, all while keeping the performance loss below 2%,” said Ilhan.

With a new grant from The Rockefeller Foundation, combined with funding from the National Science Foundation, researchers aim to create CellWatch, a technology ecosystem comprised of a mobile application for network measurements, a community planning dashboard and map, and a cellular quality prediction tool. CellWatch will enable everyday people to take connectivity measurements and merge their data with others in their community. This will allow communities to build maps of coverage and challenge cellular provider claims of coverage that are often misrepresented across Federal Communications Commission’s (FCC) maps, preventing communities from applying to available federal funding. The goal is to eventually have informed machine learning algorithms and statistical analysis be able to predict the quality of service in other areas that have not yet been measured and empower communities to improve their own cell coverage.

“Mapping mobile broadband turns out to be a very hard technical problem because quality of coverage depends on complex factors, not just how close you are to a tower,” said Ellen Zegura, a professor in the School of Computer Science.

Zegura is one of three principal investigators on the grant. Her collaborators include Elizabeth Belding, a computer science professor at the University of California, Santa Barbara, and Morgan Vigil-Hayes, an assistant professor in computer science at Northern Arizona University. At Georgia Tech, Scott Robertson in the Institute for People and Technology is leading the software development effort and Yao Xie, an associate professor in the H. Milton Stewart School of Industrial and Systems Engineering (ISyE), will be doing machine learning and statistical modeling.

Connectivity Complications

Understanding cell coverage and quality is difficult. The FCC requires all network providers report their coverage areas, but provider maps generally overestimate where coverage exists. While the FCC makes the provider coverage data available to the public, it’s notoriously inaccurate.

It’s even harder for everyday customers to figure out. To challenge a provider’s coverage claim, the FCC has a 100 page document detailing its requirements for citizen reporting including breaking down the coverage grid into hexagons and requiring citizens submit data from multiple hexagons to prove an area is lacking service. There are also requirements for the mix of positive and negative measurement results and the time of day of measurement.

“There’s a huge gap between the FCC requirements document and what regular people understand about connectivity and quality,” Zegura said. “Our goal with CellWatch is to empower everyday citizens to get involved in advocating for high quality internet to their communities to increase access to services, employment, health, and education.”

Creating CellWatch

The researchers involved intend CellWatch to make it accessible for citizens to map their own networks and challenge the FCC through CellWatch’s three-prong project:

• CellWatch Mobile Application: The Android measurement app will be built to comply with FCC requirements so users can successfully challenge provider coverage claims. A backend database will maintain data security and allow aggregation of measurements from different sources.
• CellWatch Community Coordination Tool: This interactive dashboard and map will aid citizens in organizing campaigns to challenge providers while meeting FCC requirements.
• CellWatch Prediction: Machine learning algorithms will predict mobile broadband performance using data collected by CellWatch tools and other public datasets.

All tools will be publicly available and open source to enable access for everyone. Ultimately, CellWatch’s goal is let citizens into the reporting process to democratize coverage and eventually collect enough data that the process can be automated.

“It's hard to know where there's coverage, and you're never going to measure everywhere,” Zegura said. “But if you have taken measured data in some places, you can use machine learning and statistical analysis to make predictions in new locations.”

The information won’t just aid citizens but can also be used to help allocate federal funding to add infrastructure where and when needed. The project is just the start of ensuring all of America has access to reliable network coverage and doesn’t miss out on any opportunities. Out of network will be a thing of the past.

From the movie house to the warehouse to your house, AI seems like it's everywhere. That's because with a steady stream of new research and innovative applications reaching into nearly every industry and business sector, it is everywhere. Nevertheless, AI still holds enormous potential as the field continues to evolve.

To get a sense of what this evolution could look like in 2023, we turned to a small group of Ph.D. students from the College of Computing community that are currently pushing foundational and applied AI research forward in a broad spectrum of disciplines and fields.

The students shared their opinions on where AI might be headed in the new year, what some of the big tech stories could be, and why ethics in AI are so critically important.

Where will artificial intelligence and machine learning have the most impact in 2023?

"Artificial intelligence and machine learning will continue to have a huge impact on manufacturing and warehouses with labor shortages and worker turnover continuing to be a concern as more manufacturing and operations jobs are brought back to the United States from overseas. Additionally, AI/ML will continue to help ensure that manufacturing and warehouse facilities are operating as efficiently as possible from energy and material savings to worker safety and parts quality." – Zoe Klesmith Alexander, computational science and engineering Ph.D. student

"Right now, deep learning is on a trajectory to transform the creation space. Artwork and images, videos, data representation and storytelling, co-authoring, and summarizing documents... It's gotten really good." – Ben Hoover, machine learning Ph.D. student

"I think machine learning and AI will keep playing a huge role in how the world and society will be shaped over the next decade in many ways. It will make many other fields more efficient through ML and AI tools we are developing. In 2023, I think ML and AI will have the most impact on social media platforms, helping reduce hate speech, rumor spread, etc." – Agam A. Shah, machine learning Ph.D. student

"One of the big impacts this year may be driverless cars being in your neighborhood. Otherwise, it will be a slow steady drip of GPT3 and other OpenAI models suffusing all applications, making programmers much faster, making journalists faster, making academic articles and lit reviews much faster. We're at a 4th grader level, and I hope by the end of this year we'll be at the 6th grader level. Also, indoor turn-by-turn navigation will be everywhere in 2023 as well." – Brandon Biggs, human-centered computing Ph.D. student

What will be some of the big tech stories in 2023?

"ChatGPT and the GitHub Copilot lawsuit will keep making it into the news and cause more controversies. In general, AI ethics will become more important and get more focus as the technology keeps advancing." – Fabian Fleischer, cybersecurity, and privacy Ph.D. student

"Driverless car fleets will be coming to a city near you. A new battery technology will allow phones to keep their charge for a week. Meta realizes virtual reality (VR) head-mounted displays are for a limited market and uses headphones and phones to provide VR experiences." – Brandon Biggs

What’s an issue or industry that you think could benefit from a computing solution?

"Our reinterpretation of modern deep learning as energy-based associative memories has the potential to transform any industry that relies on foundation models -- giant architectures that require models that are "self-supervised" (learn on their own from data)." – Ben Hoover

"Inclusion in everything. Over 90 percent of websites on the internet have elements that are inaccessible to 25 percent of the world's population who have disabilities. Inclusive design will be the most important area where technology can be redesigned and created to have multiple sensory modalities and be properly programmed." – Brandon Biggs

"Currently, financial markets are far from efficient because they do not fully incorporate information available in large unstructured text data. With the latest development in natural language processing techniques, we can better understand the economy and therefore price financial markets better." – Agam A. Shah

There’s been increasing recognition of the vital role ethics should play in artificial intelligence. How do you see this issue evolving in the next year?

"Specifically in my research, I think explainable AI (XAI) is very important, especially if non-experts in ML will be using black-box ML solutions in a factory. It will be important for humans to trust and to understand the models especially if the models are being using to monitor quality on a safety-critical part.

"Additionally, using XAI for human interaction with robots that utilize deep learning to make decisions will be increasingly important as technologies like collaborative robots (cobots) are integrated into factories. I think in my area of research that it is always important to use automation to aid humans in jobs that are safe for humans to do and not to replace them." – Zoe Klesmith Alexander

"Big data is pretty much at its peak. Deep data, where your Alexa knows everything about you, or your phone knows everything about you, and rather than saying 'other people who watched this show liked this show,' it's going to say, 'I know you liked these shows, I think you'll like this show because of these reasons, one of which is because other people who liked all these other shows liked this show.' The ethical element will be how much of this data should these models use, and are people going to build a personal dataset that they can share with other apps, or is each app going to need to build their own dataset? The ethical question is who owns this data." – Brandon Biggs

"I think ethics will become more and more important going forward. We are making huge breakthroughs in machine learning and artificial intelligence, but the systems we are creating are producing racist, sexist, and stereotypical results. For example, a recent system, Galactica, developed by Facebook (Meta) is powerful. It can produce research articles by just simply providing it with the title. It comes with some serious ethical concerns, in some cases, it produces racist, sexist text. So, as we will keep developing better models and making success in parallel, we need to always keep in mind the ethical implications of these models." – Agam A. Shah

What research are you working on that you think people should know about or will have impact in 2023?

"Part of my research focuses on data-driven modeling of additive manufacturing processes to better control dimensional quality of the final part. Another part of my research focuses on detecting anomalies in real-time using computer vision and machine learning for both warehouses and manufacturing processes." – Zoe Klesmith Alexander

"Right now, deep learning is built on feed-forward mathematical operations that have little resemblance to the brain. I am working on a physics inspired approach to deep learning built around recurrent networks and energy functions. These architectures have the same mathematical foundation as the famous, biologically plausible Hopfield Network." – Ben Hoover

"I am currently working on two projects which, in my opinion, will have an impact in 2023. In one project, we are measuring the exposure of public firms to ongoing inflation. We are also understanding how inflation affects different firms differently based on the pricing power of the firm. As inflation is the highest in the last 40 years, our study is highly relevant now and in the coming years till we get inflation back in control.

"The second work is related to the first work in some ways. As inflation is rising, to control the inflation Federal Reserve Bank is tightening its monetary policy. In our second work, we are measuring the stance of monetary policy (measuring hawkish vs dovish stance) of the Fed using state-of-the-art NLP models to see its impact in various financial markets (Treasury market, Stock market, Crypto market, etc.)" – Agam A. Shah

PetTrack uses a combination of sensors to give the accurate, real-time indoor location of an animal. Ultra-wideband (UWB) radio wireless sensors locate the pet and accelerometers determine if it’s sitting or moving regardless of objects or walls in the way, giving owners more detail on what their pet is doing than a camera or GPS. All of this is located on a small sensor that can be put on a collar for minimal invasiveness and can be viewed via a compatible smartphone app.

“PetTrack comprises two things: one is knowing the pet’s indoor location and second is trying to understand their activity,” said Ashutosh Dhekne, an assistant professor in the School of Computer Science (SCS).

Dhekne and his students presented the research in the paper “PetTrack: Tracking Pet Location and Activity Indoors” at BodySys 2022 in July, a workshop on body-centric computing systems that was part of MobiSys 2022 in Portland, Oregon.

How PetTrack Works

PetTrack’s innovative combination of sensors makes it unique compared to other pet-monitoring devices. The UWB radio wireless signal locates where the pet is in the home from up to 100 feet away, while the accelerometer acts as an inertial sensor that can track the pet’s pose. This means owners can learn whether their pet is standing, sitting, or even lying down.

Unlike with cameras, owners can always know where their pet is because the UWB network is accessible through walls, furniture, doors, or anything else a cat or dog can hide behind. The UWB network is plug-and-play and doesn’t interfere with existing WiFi but should be connected to a home WiFi network to give the owner smartphone updates. Location data takes up far less bandwidth than images and doesn’t burden the owner’s WiFi. Multiple UWB sensors and a central anchor data collection module help determine the location via multilateration, or individual distance measurements from different anchors,keeping it accurate.

“Together, combining where the pet is and what the orientation of the pet is, we can create a summary map of where the pet has been during the day and what activity the pet was doing,” Dhekne said.

This could reassure owners who are concerned about pets getting into forbidden places or comfort owners worried about their sick or elderly animals.

PetTrack Privacy

Privacy is another major advantage of PetTrack. Just like a router, PetTrack has at least one central anchor device in the home or backyard, making it less invasive than a camera in every room. It also doesn’t take up the entire room because the device is small enough to hide behind a lampshade . One set of anchor devices can track up to four pets with updates every second. PetTrack functions even when only a few anchor devices are reachable, albeit with lower location accuracy.

PetTrack’s local focus also keeps it safer. Whereas hackers can take over cameras and compromise privacy for owners, PetTrack’s network only functions in the area it’s in. This means it doesn’t have the same vulnerabilities cameras do such as enabling a hacker to watch owners or locate expensive objects or valuables in the house.

PetTrack’s Future Possibilities

Currently PetTrack is designed just to monitor an animal’s location and position, but it has a lot of potential. Pet daycares could use the technology, so owners have a sense of how their dog is doing away from home. It could also become a training tool where a buzzer could sound if a pet were in an area they were not meant to be.

PetTrack could not only be the future of animal-monitoring technology but could provide a whole new way for owners to connect with their dogs or cats.

“Overall, the idea is to connect better with your pet, using PetTrack,” Dhekne said. “You could detect changing pet behaviors and interact with the pet using location-aware robotic toys.”

Other researchers think PetTrack could improve the pet training process. Neeraj Alavala, the lead SCS master’s student on the project,  shared the feedback other researches provided at BodySys.

“One good piece of feedback that we got was that we could use our existing setup to also track when pets make accidents in the house during potty training,” Alavala said. “We already have the technology to track when and where these accidents happen and can ensure that those areas get cleaned up. As an extension, we can also give the pet feedback like a buzzer to train the pet to not go in the house.”

Whatever its applications, PetTrack could make the bond people have with their pets even stronger.

CITATION: Alavala, N., Chen, H., Singhal, R., Dhekne, A. PetTrack: Tracking Pet Location and Activity Indoors. BodySys (2022). https://doi.org/10.1145/3539489.3539587

The annual award, which includes a $2,500 cash prize, is one of the most prestigious mid-career awards in computer architecture and recognizes contributions to the community. Qureshi earned this year’s distinction for work he’s done to advance high- performance memory systems since earning his Ph.D. in 2007.

“This is a wonderful and much deserved honor - my heartiest congratulations,” said Babak Falsafi, EPFL computer science professor and SIGARCH chair.

The announcement was made during an award luncheon during ISCA'22: The 49th International Symposium on Computer Architecture, which was held in New York City this summer.

“Research is a very rewarding profession. When you solve a difficult problem, the kick that you get is completely unlike any other. When your ideas get adopted, the satisfaction you get is unlike any other.

“And if the community recognizes the work, that’s the cherry on top,” said Qureshi during remarks at the event.

The award is named after Maurice V. Wilkes, a British computer scientist and a Turing Award winner. Wilkes is widely considered as the father of microprogramming and played a pivotal role in the electronic delay storage automatic calculator, which started operating in 1949.

Qureshi's research has impacted many design aspects of high-performance memory systems according to School of Computer Science Chair Vivek Sarkar.

“Moin's selection for this award clearly demonstrates the leadership impact of his research career thus far and reflects positively on both his accomplishments and on Georgia Tech.

“His highly cited publications have become household names in the computer architecture community and continue to have a major impact on subsequent research and product development,” said Sarkar, who is also the Stephen Fleming Chair for Telecommunications in the College of Computing.

[RELATED: Professor Wins Intel Outstanding Researcher Award]

Sarkar points to two facets of high-performance memory systems design– caches and non-volatile memory (NVM) – as examples of Qureshi’s lasting contributions to the field.

He took a new approach to the design and configuration of memory caches by introducing novel policies. These policies placed larger caches at the outermost level of the cache hierarchy relative to smaller caches placed at the innermost level. This new approach ultimately influenced how cache architectures for 3D memory systems are developed.

Qureshi’s work has also influenced the evolution of NVM and the design of future hybrid memory systems. His research demonstrated that phase-change memory can be combined with standard dynamic random-access memory to create a hybrid system that delivers the benefits of both memory types. This breakthrough allowed for cost-effective and power-efficient increases in memory capacity.

“We are all proud of this achievement and can already see signs of influential impact of his more recent research in computer architecture, including in the field of quantum computing,” said Sarkar.

Qureshi, who holds nearly two dozen patents, began his career as a research staff member at the IBM T.J. Watson Research Center, where he developed the caching algorithms for Power-7 processors.

He earned best paper award in 2018 at  IEEE/ACM International Symposium on Microarchitecture,. Qureshi won a 2019 Persistent Impact Prize for his ISCA 2009 paper on phase-change memory and 2021 Persistent Impact Prize for his MICRO 2009 paper on PCM wear leveling, both recognized for “exceptional impact on the fields of study related to non-volatile memories.”

Along with his research, Qureshi is contributing to the field by developing the next generation of researchers, with several of his Ph.D. students placed at top academic places, including at UIUC, University of Wisconsin, University of British Columbia, and University of Toronto.

He is the director of the Memory Systems Lab, which designs and develops future computing systems. Current work in the lab focuses on quantum computing, secure hardware, and robust artificial intelligence.

UWB technology uses low-level energy signals to enable short-range, high-bandwidth communications that do not interfere with other types of radio communications.

To push forward on what’s next in the field, School of Computer Science Assistant Professor Ashutosh Dhekne recently received a National Science Foundation (NSF) CAREER award for his proposal, Closing the Gaps in UWB Localization and Sensing: Algorithms, Architectures, and Prototypes.

Dhekne expects his UWB research to not only eliminate the need for cameras to sense and track the minute movements of a robotic arm, he anticipates that it will also enable the development of infinitely scalable indoor navigation systems and advanced intrusion detection systems.

“The overarching goal of this proposal is to carefully address some of the important challenges for indoor localization and sensing using wireless signals. This award provides an impetus to advance the state-of-the-art for indoor localization and will help support my research in the field,” said Dhekne.

The challenges Dhekne mentions relate to current algorithms and system architecture.

The technology shows promise because UWB frequencies can easily penetrate walls and the returning signals can be precisely timed. However, the algorithms and architecture used for current wireless systems cannot meet the needs for large indoor spaces.

“When we try to use the same algorithm­s for fine-grained localization, such as for localizing a pen on a whiteboard or tracking a robotic arm, we face hard-to-overcome limits of precision,” said Dhekne.

Overcoming these limitations won’t be easy. But Dhekne is confident he and his team of student researchers can develop innovative architectural solutions to dramatically advance the capabilities of UWB technology. These potential solutions include developing new multi-antenna systems that leverage the advantages of UWB and use signal phase to pinpoint an object’s location to within a millimeter.

Once more robust algorithms and improved architectural structures are developed, the team plans to move from the theoretical to the practical and create working prototypes to demonstrate the feasibility of their work.

[RELATED: Newest Professor Explores the Possibilities of Wireless Sensing]

The expected improvements in fine-grained location and precision tracking of the 3D movements of a robotic arm will enable systems to detect deviations from nominal behavior. This is crucial to improving maintenance, auditing, and compliance record-keeping efficiencies for automated devices.

Coupled with these advances in sensing and tracking, the ability to infinitely scale systems will enable next-generation personal navigation in large spaces.

Using UWB receivers that capture signals sent by installed anchor devices, these systems will allow an unlimited number of users to pinpoint their exact location and navigate airports, parking decks, and other complex indoor areas. These systems will ensure privacy because they do not require a user’s device to transmit a UWB signal.

Advances in UWB sensing and tracking capabilities could also be used to provide better context to digital personal assistants.

“Localization makes digital personal assistants such as Google Home and Alexa more context aware. Commands given to digital assistants can become more natural when the user’s location is known.

“For example, just saying ‘turn on the lights’ can be understood as a request to turn on the living-room lights if the user is in the living room,” said Dhekne.

Another application of wireless sensing, according to Dhekne, will be for advanced security systems. Transmitted UWB signals travel in all directions and bounce off obstructions in an indoor space. People entering a space will cause detectable disturbances of the wireless signal-reflection patterns.

This approach allows a UWB-based system to ignore disturbances caused by a pet, an elderly person living in the house, or another known entity. This capability means a security system can remain constantly armed with much less chance for false alarms.

“Along with household use, this technology can also be used in museums or other high-value locations where guards are also present, since the proposed idea can ignore the movements of friendly entities,” Dhekne explained.

In addition to supporting Dhekne’s theoretical and practical research, his NSF CAREER award is also funding several upcoming educational activities related to the work including graduate student projects focused on mobile computing and the internet of things (IoT). These projects will be part of Dhekne’s CS8803-MCI course offered this coming Fall.

Educational activities planned through this grant go beyond just Georgia Tech students. For example, Dhekne and his team plan to create educational IoT kits and conduct workshops for students from local middle and high schools.

The team also intends to open source the software and other technologies produced from their research. This will enable a larger community of researchers and enthusiasts to explore wireless localization and sensing.

As for what’s next, Dhekne says that a lot more can be done to advance UWB technologies.

“In the future, I hope to work on innovative solutions to some of the most pressing issues that we face today such as sustainability.

"Wireless localization and sensing can reduce resource waste through smarter, contextually aware IoT in homes, factories, warehouses, and other large indoor spaces. Localization can improve efficiencies of these vast spaces, help us find lost items, and overall improve quality of life,” said Dhekne.

“Scientific research involves increasingly complex software, technologies, and platforms,” said Alessandro Orso, the software engineer and professor of computer science who is heading up the project. “Also, platforms constantly evolve, and the complexity and amount of data involved is ever-growing.”

The result is that these software systems are often developed as prototypes that are difficult to understand, maintain, and use, which limits their efficacy and ultimately hinders scientific progress.

Software engineers are trained to address these kinds of issues and know how to build high-quality software, but their time is too expensive for a typical research project’s budget. In typical grants, software is often treated as a byproduct of research, meaning that limited funding is allocated for it.

That’s where Schmidt Futures comes in. Schmidt Futures is a philanthropic initiative founded by Eric and Wendy Schmidt that bets early on exceptional people making the world better. They are investing $40 million in VISS over five years at four universities: Georgia Tech, University of Washington, Johns Hopkins University, and University of Cambridge.

“Schmidt Futures’ Virtual Institute for Scientific Software is a core part of our efforts to mobilize exceptional talent to solve specific hard problems in science and society,” said Executive Vice President Elizabeth Young-McNally.

At Georgia Tech, the funds will hire a software engineering lead, as well as three senior and two junior software engineers. A faculty director and an advisory board will help guide the group’s work, which will include collaborations with Georgia Tech scientists.

"We are very proud to host one of the four inaugural Schmidt Futures Virtual Institute of Scientific Software centers,” said Charles Isbell, Dean and John P. Imlay Jr. Chair of Computing.

“Georgia Tech’s center will advance and support scientific research by applying modern software engineering practices, cutting-edge technologies, and modern tools to the development of scientific software. The center will also engage with students and researchers to train the next generation of software engineering leaders.”

“I am excited to receive the award,” said Zhang, an assistant professor in the School of Computer Science (SCS). “As a programming languages researcher, I enjoy what I do and believe it is important to continue working on fundamental problems to improve software reliability.”

He will use the award for interleaved Dyck-reachability research, which makes program analysis simpler and more efficient by eliminating ineffective parts of the graph. This funding will lead to practical tools that can analyze real-world software and foster collaboration with Amazon automated reasoning researchers.

This work has been widely recognized by the research community. Zhang and his students have won a 2020 distinguished paper award at Programming Language Design and Implementation (PLDI), an online conference, for this work. His student, Yuanbo Li, also received a Facebook Fellowship for related research in May.

“Qirun is a world-class leader in programming languages,” SCS Chair Vivek Sarkar said. “It is great to see him receive this opportunity to increase the industrial impact of his research.”

“Georgia Tech had a very strong presence at MICRO, representing about 5 percent of the papers appearing at the conference and placing us among the top five institutions in paper count this year,” said SCS Assistant Professor Alexandros Daglis.

SCS papers included:

COSPlay: Leveraging Task-Level Parallelism for High-Throughput Synchronous Persistence

Marina Vemmou, Alexandros Daglis

Cerebros: Evading the RPC Tax in Datacenters

Arash Pourhabibi, Mark Sutherland (EPFL); Alexandros Daglis (GT); Babak Falsafi (EPFL)

Vortex: Extending the RISC-V ISA for GPGPU and 3D-Graphics Research

Blaise Tine, Krishna Praveen Yalamarthy, Fares Elsabbagh, Kim Hyesoon

JigSaw: Boosting Fidelity of NISQ Programs via Measurement Subsetting

Poulami Das (GT); Swamit Tannu (University of Wisconsin-Madison); Moinuddin Qureshi (GT)

ADAPT: Mitigating Idling Errors in Qubits via Adaptive Dynamical Decoupling

Poulami Das (GT); Swamit Tannu (University of Wisconsin-Madison); Siddharth Dangwal (IIT Delhi); Moinuddin Qureshi (GT)

“This is both an exciting and disruptive time for computer architecture research as we approach the end of Moore’s law,” said SCS Chair Vivek Sarkar. “It’s great to see this forward-looking research by Georgia Tech faculty and their collaborators being showcased at MICRO.”

The tool uses wireless location, inertial sensors, and pressure sensing to determine where the pen is on the board and digitize those movements to stream to a computer. Costing only $100-$150, it’s nearly a tenth of the price of a smartboard. For an hour lecture, it uses just 1.5 MB of data, which is about 400 times smaller than filming the whole board.

Before this innovation, the two primary ways of digitizing content both had logistical problems. While a smart board is accurate, it’s often prohibitively expensive for most school budgets. Although filming video of the whiteboard is more cost effective, it results in large video files that don’t transfer well over low internet speeds. The potential inequity of this among students inspired the researchers to work on solutions.

“Not being able to digitize the whiteboard increases the learning gap,” said School of Computer Science Assistant Professor Ashutosh Dhekne. “If you have good internet connection, you can see the whiteboard clearly in a video, but if you don’t, the learning experience deteriorates quickly.”

The researchers focused on making the experience seamless for the teacher. So, they decided not to modify the whiteboard or tip of pen itself but focus on attachments.

The attachment needs to sense where the pen was as a person wrote. Location tracking alone isn’t sufficient, so the researchers relied on a fusion sensor approach:

• Inertial sensors: Each sensor contains an accelerometer, gyroscope, and magnetometer that can determine exact orientation of the pen on a larger frame of reference, in this case the whiteboard.
• Wireless radio technology: This tracks how far a pen is from the fixed wireless antenna sensors at the border of the whiteboard.
• Pressure sensor: This sensor detects when the pen touches the whiteboard surface, based on how much pressure is coming from the user’s grip.

The researchers believe this attachment could be used for more than just online learning and meetings; it could be also attached to scalpels or paintbrushes for medical or art instruction.

“Documenting the process of creating something by hand is difficult, but this tool has the potential to make that possible,” Dhekne said.

Dhekne, SCS Professor Mostafa Ammar, and SCS Ph.D. student Yifeng Cao presented the research in the paper, ITrackU: Tracking a Pen-like Instrument via UWB-IMU Fusion, at The 19th ACM International Conference on Mobile Systems, Applications, and Services  (MobiSys) in June.

Twenty students submitted applications for the fellowship and four were selected by a committee of SCS faculty members including Professors Dana Randall and Ellen Zegura, Associate Professor Ada Gavrilovska, Assistant Professor Qirun Zhang, and Senior Research Scientist Greg Eisenhauer.

The fellowship recipients, along with their primary advisors and faculty collaborators, are:

• Alind Khare (Assistant Professor Alexey Tumanov, with Associate Professor Tushar Krishna)
• Xinya Liu (Professor Alessandro Orso, with Assistant Professor Joy Arulraj)
• Marina Vemmou (Assistant Professor Alexandros Daglis, with Professor Moin Qureshi)
• Renzhi Wu (Assistant Professor Xu Chu, with Assistant Professor Joy Arulraj)

“I think the SCS Incubator Fellowship is a great initiative to kickstart collaborations,” said Daglis, who will be revisiting research on memory hierarchy design policies for modern architectures with intense networking activity. “It is exciting to embark on this joint SCS-backed research effort together, combining our expertise in memory systems and networked architectures.”

Winners will have the 2021-22 academic year to seed their collaborative research and present their work at the end of the academic year.

“I am delighted by the outcome of this first round of the new SCS Incubator Graduate Fellowship Program,” said SCS Chair Vivek Sarkar. “It’s exciting to see so many junior faculty members involved, and I look forward to seeing the fruits of these new collaborations.”

The $150,000 planning grant will enable researchers to explore the possibilities of the project, develop community relationships, and gather a team of students. They believe their methods in the Westside could serve as a test case for applying the solution to similar urban neighborhoods elsewhere.

“The digital divide in urban communities is getting worse: They don’t have reliable connectivity, or affordability for connections,” School of Computer Science (SCS) Professor Umakishore Ramachandran said. “Georgia Tech should be able to do something to help our neighbors.”

The project started three years ago when researchers approached Westside leaders to figure out their needs for increasing access to information. To develop trust with the community, undergraduate and graduate students have been regularly attending Westside community meetings.

They determined the community would most benefit from local area network (LAN) based connectivity and streaming services that are not reliant on cloud connectivity. For sustainability, they aim to train high school students and other community members to maintain these systems.

Building relationships is just as important as the technology. As a concrete use case, they plan to provide a streaming service using an open-source LAN-based platform akin to Zoom for Tech students to tutor Westside area children.

“Technology is the easy part; the hard part is earning trust of community,” Ramachandran said.

Ramachandran is co-principal investigator along with SCS Assistant Professor faculty Ashutosh Dhekne and Professors Mustaque Ahamad, and Mostafa Ammar; School of Economics Associate Professor Shatakshee Dhongde; and City and Regional Planning Associate Professor Nisha Botchwey.

After the faculty held a graduate student listening session in December 2020, it was clear students wanted more peer support.

To this end, a group of graduate students started creating programing to connect to their fellow students. The first event was a game night and photo contest judged by Professor Alexandra Boldyreva held on Gathertown, a participatory video chat platform that creates a virtual party environment. Next they hosted a panel called "Things nobody told me about a Ph.D.!"

With the success of early programming, the group formed and officially registered as the first-ever School of Computer Science Graduate Student Association (SCS GSA), which officially launched in August.

“It’s great to see our students take the initiative to come together and form this organization,” SCS Chair Vivek Sarkar said. “I look forward to the SCS GSA contributing to an increased vitality in our school’s community.”

The founders of the association say they hope to create a sense of belonging for students, improve student – faculty relationships, and provide students with an opportunity to give back and become leaders.

“I see it as an avenue to create cherished memories of our graduate school journeys beyond the walls of our research and labs,” said Harshit Daga, SCS GSA president and final-year Ph.D. student.

With this in mind,  SCS GSA’s programing is tailored to be both practical and personable. The group recently helped the school with onboarding the incoming master’s and Ph.D. students by combining them on a single Slack platform. For the semester the team has planned events such as a Ph.D. welcoming event, monthly TGIF gatherings, and townhalls.

“I envision GSA to better the lives of graduate students by improving the mental health and well-being of the student body and providing opportunities for them to succeed,” said third-year Ph.D. student and treasurer Sharjeel Khan.

Along with Daga and Khan, the GSA leadership team includes: fourth-year Ph.D. student Bhuvesh Kumar as vice president, second-year Ph.D. student Manasvini Sethuraman as events chair, second-year Ph.D. student Mehrdad Ghadiri as student and faculty affairs chair, and third-year Ph.D. student Kuchhal as communications chair. SCS alumna Thaleia Dimitra Doudali (CS Ph.D. 21) serves as alumni liaison.

“I really wanted to stay in academia so I could work on cutting-edge research and problems I am excited about, and lead my own research group,” Doudali said

Doudali will build from her research on memory systems, using machine learning to optimize resource management. IMDEA is a world-renowned research institution that builds cost-effective and high-quality software.

The Spanish location also lets Doudali establish herself in the European research community she started in when she completed her undergraduate education at the National Technical University of Athens (NTUA).

“I am not just going to be a professor but part of a community, collaborating with researchers and industry partners, attending conferences, reviewing papers, and having even more exposure to cutting-edge ideas,” she said.

Thriving at Tech

Research collaborations have always been important to Doudali, who chose SCS for her Ph.D. for this reason.

“Georgia Tech is a great place to do research because it’s big, very diverse, and you have a lot of support,” she said.

A big part of the support Doudali had was her advisor, SCS Associate Professor Ada Gavrilovska, who introduced her to heterogeneous memory systems research. This area builds novel systems that use new memory hardware technologies with maximum application and resource efficiency.

Doudali’s dissertation contributed system-level mechanisms to manage heterogeneous memory resources with machine learning methods. It provided the design foundations for the practical use of machine learning and lays the grounds for future use of machine learning and computer vision methods for resource management.

Her work has received numerous accolades: a best paper award finalist at the 28th International Symposium on High-Performance Parallel and Distributed Computing (HPDC 2019) and a Rising Star Award for Doudali.

Along with guiding her research, Gavrilovska also advised on Doudali’s career track. Although Doudali did several industry internships, including VMWare, AMD Research, and Dell EMC, Gavrilovska was very supportive and encouraging during Doudali’s academic job search.

“Having a female role model has helped me immensely,” Doudali said. “She was always very good at letting me improve my technical and soft skills and encouraging of service work. It’s very important to work with someone who is not only a great researcher, but a lifelong mentor.”

A Legacy of Service

Doudali hopes to pay it forward as a professor herself by mentoring students. Service has been an important part of Doudali’s career at Georgia Tech. In fact because her Ph.D. visit day convinced her to attend SCS, and she volunteered on visit day every year during her time at Tech.

“When you have a nice experience, you want to recreate it for future generations,” Doudali said. “It was one of the early signs of me wanting to go academia because I really like to mentor students and give advice about Tech and Atlanta.”

After discovering this passion, Doudali decided to be more involved than just one day a year. She helped found the SCS Graduate Student Association to advocate for and bring the school’s graduate students together. Although Doudali will be graduating before the group can be official, she will help with alumni relations. 

Doudali also held multiple roles at Georgia Tech’s Hellenic Society for Greek students. Throughout her years on campus, she has helped organize events to bring together students and professors of Greek heritage.

She credits this service with helping give her the soft skills that made her excel in the academic job market: communication, organization, and collaboration.

“Even if you want to focus on research, it’s always useful to talk to people and socialize to see what other people are working on and having another opinion on your work, while making friends and enjoying your free time,” she said.

Doudali is excited to bring everything she learned at Tech to her new role.

“I am excited for this new chapter in life and I will always cherish the great friends and moments I had while at Georgia Tech.”

The new site will provide a faster and more user-friendly digital experience while offering greater accommodation for those with accessibility needs.

Two major components of this redesign includes menus organized by user groups rather than by departmental structure and an enhanced mobile-friendly experience that is compatible with all major browsers and across all devices. 

This new streamlined experience is further complemented by the updated Georgia Tech branding theme which features the iconic Tech Gold header and characteristic Institute-wide slogan, Creating the Next.

Users will notice a number of other engaging new features. These range from the ability to sort faculty members by school to being able to sort events by type and function. 

The research needed to complete such a major transformation was compiled by two student teams from the School of Interactive Computing master’s in Human Computer Interaction (HCI) program. 

Team members Harshali Wadge, Santiago Arconada Alvarez, Prabodh Sakhardande, Shihui Ruan, Jordan Hill, Jordan Cox, Chaoyuan Luo, Yuhan Zhou, and Lu Meng were all leads on this research as part of the HCI Special Topics class taught by Senior Research Scientist Carrie Bruce. 

These student teams spent the 2019 Fall semester assembling a series of evidence-based design methods, field surveys, and testing groups that were used to inform the overall user experience of this new site.

With the student teams’ initial research and the efforts of a dedicated team of staff, the college has successfully condensed several thousand pages of content into hundreds. This aggregation and purging of old content will allow all audiences to enjoy a more up-to-date and direct experience.

For any questions, suggestions, or updates upon launch, please complete the Website Feedback Form which is located on the main menu under the About dropdown tab.

To help students begin thinking about ethics with regards to research, faculty at Georgia Tech – in conjunction with Mozilla – held the first workshop on integrating ethics and responsible computing into courses this summer.

The workshop was a collaboration between faculty researchers at Georgia Tech in both the Ethics, Technology, and Human Interaction Center (ETHICx) and Computing and Society, as well as Mozilla. The workshop received a strong response, which organizers say indicates a growing desire for ethics at the center of computer science courses.

Members of the College of Computing’s Division of Computing Instruction, the Schools of Interactive Computing, Computational Science and Engineering, Computer Science, and Electrical and Computer Engineering, along with attendees from Georgia State all participated in the online workshop.

“It’s really gratifying to have broad representation because it demonstrates the desire for people from so many different areas to think more deeply about the role of ethics in our education,” said Ellen Zegura, professor in the School of Computer Science and Fleming Chair in Telecommunications.

The goal of the workshop was to help instructors consider ways in which to implement ethics as a central piece in courses not just later in a student’s study, but from the very beginning. There’s an issue of urgency, Zegura said, that needed to be considered.

“Computing has reached a point where it is being used for critical decision making that really affects people’s lives,” she said. “The need to use computing responsibly has moved up incredibly. And if we don’t talk about ethics early in the curriculum, we’re sending a message that it’s not important. If you only hear about it in one course and it’s later in your career, then what does that say about the importance? Students see that.”

While official plans aren’t currently in place to continue the program, Zegura said the idea is to continue this as a series of activities that are responsive to what people’s needs are, specifically those who want to do a better job of embedding ethics into their computer science curriculum.

Georgia Tech graduate Kathy Pham (CS ’07, MS CS ’09), now at Mozilla, has been instrumental in engaging the computer science community from 15-20 universities on focusing on ethics, Zegura said.

Portions of the workshop can be viewed on YouTube here.

School of Computer Science (SCS) alumnus Ashish Bijlani discovered this at the start of his Ph.D. when he tried to install an operating system update and was told he didn’t have enough space on his device.

“I don’t play a lot of games, take a lot of pictures, and have very few apps I installed myself, so I was shocked. I wanted to find out what was going on so I started to look into this,” he said.

Bijlani spent his Ph.D. studying phone storage consumption to confirm how much space apps were really taking up and how little users were actually engaging with them. His solution is to make context-specific storage that follows the behavior of the user.

Studying Storage

To determine how each app uses storage, Bijlani had to download the most popular apps and analyze their consumption. Downloading millions of apps to determine their storage size required a large distributed system. When Bijlani started the project in 2014, he downloaded 1.1 million apps; by 2016 the number was up to 2.2 million.

The team analyzed popular apps released in two-year intervals from 2014 to 2019. When an app becomes especially popular, it often turns into a super app with more features than its initial purpose. For example, Uber may have started as a ride-sharing app, but now offers food delivery and scooter rental.

“Developers add a lot of features so users stay engaged with the app because that’s how they monetize it,” Bijlani said. “But only a few users are going to use the majority of these features.”

Bijlani found that most users engage with just 10 percent of app after conducting a user study with 140 participants from the University of Buffalo’s phone lab testbed, where researchers can deploy Android changes to participants’ phones and carry out studies.

Since these apps are optimized for performance, storage often gets ignored to the detriment of the user. For example, apps include residual files, such as logs on how users interact with ads, that never get cleaned up. There is no incentive for vendors to maximize storage because they can sell bigger phones or more cloud storage.

Proposing a Solution

The phone lab also showed that though most users take advantage of just 10 percent of the app, they’re not all using the same percentage. With gaming apps, for example, some users are content to play just the first few levels, whereas others play the game to completion.

Bijlani believes apps could be context sensitive to users’ behavior in the same way they already track screentime and Bluetooth devices. With that in mind, he developed a lightweight storage-tracing tool, Cosmos, that collects traces of how users interact with apps to determine the best customized storage model. For example, if a user only opens the weather app in the morning, then future phones could offload it for the rest of the day.

He presented this research at ACM SIGMETRICS in June. Bijlani co-wrote the paper, Where did my 256 GB go? A Measurement Analysis of Storage Consumption on Smart Mobile Devices , with SCS Professor Umakishore Ramachandran, and University of Illinois Urbana-Champaign’s Professor Roy Campbell.

“One of the bittersweet joys of being a faculty member is seeing your students graduate,” said Chair Vivek Sarkar.  “The bond between a graduate student and their advisor is lifelong. I wish all the graduating students the best of success and happiness in the future.”

The Ph.D. students included Prithayan Barua, Bobin Deng, Ren Ding, Prasun Gera, Ramyad Hadidi, Jinho Jung, Joshua Kimball, Xiangyu Li, Stacey Truex, and Jun-Kun Wang. The master’s students were Jonathan Ciak, Apruv Singh Gautam, Andrew Leitheiser, Maneel Modi, Tanishq Rupaal, Rahul Saranjame, Farzam Tafreshian, Ruhao Tang, Kai Yi, and Yiwen Zhou.

The students are off to exciting futures. Gautam, who was advised by Professor Mustaque Ahamad, will be a threat researcher at Cyble after graduation. His experience as a teaching assistant prepared him for the role.

“My favorite memory is TAing CS 4235/6035 Intro to Info Security for three semesters” Guatam said. “I really enjoyed helping students and also got to learn a lot from that experience.”

Saranjame, who also was advised by and TA’d for Ahamad, will be a security engineer at Praetorian.

Barua, who was advised by Sarkar, will join SiFive, an open-source hardware startup.

Four master's students and 11 Ph.D. students graduated earlier in the academic year and are headed off to academia (EPFL, UCLA) and industry (Facebook, IBM Research).

Yet all the graduate students will miss the SCS community in different ways. Hadidi, who studied computer architecture under Professor Hyesoon Kim, will remember how supportive SCS was on every level, including catered events.

 “I could go back to lab without worrying about what to eat when hunger strikes,” he said.

CompOFA is an algorithm that trains hundreds of models simultaneously and makes this process inexpensive by identifying and focusing on the most efficient possible models.

“This research is definitely in the spirit of democratizing ML,” School of Computer Science (SCS) Assistant Professor Alexey Tumanov said. “Only large companies with resources can afford to do research like this. Usually, for state of the art accuracy, you need money, a lot of GPUs at your disposal, or a lot of time.”

Challenges with Training Multiple Models

Traditionally, developers have to design and train ML models for each platform on which they want to deploy their application. This is a slow, costly process, requiring ML experts to put in several days’ worth of computation on expensive hardware. Techniques like Neural Architecture Search (NAS) aim to automatically find good ML models but are even more resource intensive.

More recently, weight-sharing training algorithms were proposed to produce trillions of models during a query. However, only some models in this large search space can be “good” – the vast majority of remaining models are inefficient and thus waste computation.

The researchers realized it’s possible to extract just the optimal models from this search space,  giving the most accurate models at a given latency target.

“Our observation is it’s not really necessary for specialists to spend time, resources, and effort to train this many architecture candidates if there is a high probability they won’t be optimal,” Tumanov said. “If we reduce the suboptimal space, we can come up with a faster training procedure.”

Their second key observation is that the design space, or a collection of hundreds of ML models, doesn’t need to be as dense. Rather, the researchers only need to pick the models that are sufficiently different in their size —any smaller differences are too fine-grained to be distinguishable — a systems insight leveraged to further reduce computation.

Pruning the Architecture

Once optimal models are identified and extracted, the architecture space is pruned. This makes the training speed 2x faster and search speed 200x faster. What originally would take four hours now takes 70 seconds. CompOFA also halves the cost and CO₂ emissions over previous state-of-the-art methods.

“You can remove models safely while keeping performance objectives,” SCS master’s student Manas Sahni said. “We receive results that are just as good at half the time to train the model.”

CompOFA can produce a family of models. While a lot of research in this area focuses on searching for and training a single architecture, the researchers wanted to produce a family of models that can run simultaneously, saving costs.

The researchers presented at the International Conference on Learning Representations. Tumanov and Sahni wrote the paper, CompOFA – Compound Once-For-All Networks for Faster Multi-Platform Deployment, with SCS master’s student Shreya Varshini and SCS Ph.D. student Alind Khare.

Created to fund research in top computing areas, the fellowship provides tuition support for two years and a $42,000 stipend. It also offers winners the opportunity to visit Facebook’s campus for a Fellowship Summit.

“I am very honored and excited to win the fellowship,” Li said. “I am very grateful for the help from my advisor and collaborators as well.”

Li will use the fellowship to explore better static analysis techniques for concurrent programs. These programs execute several computations simultaneously. As Moore’s law ends, concurrent computation could be a solution. However, detecting bugs is challenging even for experienced software developers.

“I am delighted to hear that Yuanbo had been selected for the Facebook Fellowship award,” said Li’s advisor, SCS Assistant Professor Qirun Zhang. “It is a highly competitive program that well recognizes his past accomplishments and future potential.”

“My research highly aligns with the scope of the fellowship on novel computing paradigms,” Noor said. “It will also help me to reach a broader audience by presenting my work at the CRNCH Summit.”

During the fellowship period, she will work on plasmonic computing system. Surface plasmon is an electromagnetic wave that merges the miniaturization potential of electron and high-speed attributes of photon. Researchers can use surface plasmon to design logic devices and integrated circuits that offer high-speed and high throughput computation with low footprint requirement. Plasmonic computing systems can be used in signal and data-processing applications in high-end server systems.

Noor’s project focuses on the design and optimization of the building blocks of a plasmonic computing network. As part of the project, she will design the couplers between plasmonic metal-insulator-metal (MIM) and metal-semiconductor-metal waveguides.

She will also explore the material choice for plasmonic MIM waveguide-based devices to determine best material factoring in energy efficiency, footprint, and speed.

CRNCH launched the fellowship program in fall of 2020 to support innovative student research in post-Moore computing topics. The research center explores new computing paradigms after the end of Moore’s law.  Partnering with academics and industry, CRNCH researchers full-stack solutions on everything from quantum computing to approximate computation.

Ensemble of Diverse Models (EDM) diversifies the models in the cloud to create discontinuous predictions and severely weakens the attacker’s ability to clone a targeted model. The defense seamlessly works with other existing defenses, does not degrade accuracy, and only involves modest computational overheads.

“People generally make changes to a single model to prevent stealing, but we are proposing a different solution of using an ensemble of diverse models that makes model stealing harder to do,” said School of Electrical and Computer Engineering Ph.D. student Sanjay Kariyappa, who is advised by School of Computer Science Professor Moin Qureshi.

How Cloning Works

In several applications, model data is very inaccessible to attackers. The easiest way for them to steal secure data is to use out of distribution data (OOD) to query the target model.

“An attacker does not have the in-distribution queries, but to clone models they do not need in-distribution queries, but OOD queries make it easier to figure out what function is and steal the model,” said Qureshi.

With the target model’s predictions, attackers can train a clone model. This clone model learns to approximate the target model’s decision boundary and can achieve a high level of accuracy on in-distribution data. 

How EDM Changes Things

EDM, however, is made up of more than just one model. This research proposes a novel training technique to create an ensemble of diverse models that are harder to clone.

Given an input query, a hashing algorithm will select a different model to provide an output. Each EDM model is trained to produce dissimilar predictions for OOD queries. In effect, these models provide highly discontinuous predictions for the attacker’s OOD queries. These discontinuous predictions make it almost impossible for an attacker to build a high-accuracy clone model.

Although the researchers only used five diverse models, the number of models in the ensemble can be scaled up to improve security. When evaluated on several image classification tasks, EDM weakens the accuracy of clone models up to almost 40 percent.

The researchers will present at the International Conference on Learning Representations. Kariyappa and Qureshi wrote the paper, Protecting DNNs from Theft using an Ensemble of Diverse Models, with University of Michigan Professor Atul Prakash.

This is the second cache attack paper for School of Computer Science Professor Moin Qureshi’s group, who have been working on secure cache architectures for the past three years.

“It helps to think like an attacker,” said School of Electrical and Computer Engineering Ph.D. student Gururaj Saileshwar , the lead author of the paper. “It is important to improve our understanding of attacks before a real attacker in the wild does so. In the process, we came up with the Streamline attack that is faster than all existing attacks and has fewer requirements.”

“Better attacks motivate better defenses,” Qureshi said. “Advancing the attack enables us to come up with good defenses for making cache memories secure.”

How Covert Channel Attacks Work

In this type of attack, attackers use a covert channel to communicate and transmit data without detection. Memory caches are susceptible because they are often shared between processors. Such channels have become more popular recently after they were used to transmit data in speculative execution attacks like Spectre and Meltdown.

Memory cache covert channel attacks take advantage of the time difference between access to processor caches and DRAM memory. Senders can influence whether a shared address is in the cache and manipulate the receiver’s access to it. The two fastest attacks have been the Flush+Reload and the Flush+Flush.

In a Flush+Reload, a sender installs an address in a cache and a receiver uses cache flush instructions to evict a shared address. In a Flush+Flush, a sender installs an address in the cache then the receiver measures the latency of the flush to access this address.

A major disadvantage of this type of attack is that it requires access to cache flush instructions, which are disabled in many new CPUs. Also, bit-by-bit synchronization between the sender and receiver that considerably slows the attack. This has limited the bit rate of current attacks to 500-600 KB/s for more than a decade.

How Streamline Works

Instead Streamline relies on asynchronous communication and makes the following improvements:

1. Streamline communicates over a sequence of shared addresses that enables the sender to keep transmitting successive bits without waiting for the receiver.
2. The addresses are preserved until the receiver can access them.
3. When the receiver accesses the address, they get evicted from the cache automatically due to cache-thrashing, the act of accessing a large sequence of addressees by the sender and receiver, without relying on flushing.

The researchers tested Streamline on an Intel Skylake central processing unit and achieved a bit-rate of 1801 kilobytes/second, which is 3.1 times faster than the previous fastest attack. Given that Streamline relies on generic cache properties, it works on all architectures.

Saileshwar and Qureshi wrote the paper, Streamline: A Fast, Flushless Cache Covert-Channel Attack by Enabling Asynchronous Collusion, with University of Illinois—Urbana Champaign Assistant Professor Christopher Fletcher. The researchers will present at the premiere systems conference Architectural Support for Programming Languages and Operating Systems (​ASPLOS) from April 12-23.

“It is a great honor to receive this award,” Qureshi said. “I have always enjoyed collaborating with Intel, right from my younger days of summer internships till now. Intel has a great set of researchers and you get feedback that keeps the research practical and useful.”

Qureshi was recognized for designing efficient and robust hybrid memory architectures. These memory systems combine conventional DRAM memory modules with emerging memory technologies such as non-volatile memory (NVM) and high-bandwidth memory (HBM).

In hybrid memory systems, placement of data can determine performance and energy efficiency. Ideally, frequently accessed data is placed in the low-latency high-bandwidth memory, and the remaining data in high-capacity low-cost memory.

Qureshi’s group has developed innovative designs such as Alloy Cache, which outperforms combines the tag and data together to do the cache lookup in a single access and outperforms most designs.

His group has also developed low-cost set-associative DRAM caches (called ACCORD), which maintain the low-latency of direct-mapped caches while providing conflict-miss reduction of set-associative caches. His group is currently investigating low-cost compression designs that are suitable for large giga-scale memory systems while limiting the performance degradation from metadata lookups.

Intel has supported Qureshi’s group for developing hybrid memory designs for more than five years. This research has produced several publications at flagship architecture conferences, such as the International Symposium on Computer Architecture (ISCA) and International Symposium on Microarchitecture. The grant has supported the Ph.D. dissertations of three graduate students.

“Intel is one of the few companies that has an accessible funding model focused on fostering academic research,” Qureshi said. “I am thankful to Intel for the continued support and look forward to the collaboration.”

This is Qureshi’s second Intel award. He first won the inaugural Intel Early Career Faculty Honor Award in 2012.

Qureshi leads the Memory Systems Lab at Georgia Tech. His research group looks at a variety of topics ranging from quantum computing to hardware security to robust artificial intelligence algorithms to designing future memory systems.

This IEEE Board of Directors honors senior members of the organization who have advanced science, technology, and engineering for the benefit of society.

Lee said it was “truly an honor” to be recognized for contributions to machine-learning-based approaches to intrusion and botnet detections.

Orso was recognized for his contributions to software testing and debugging.

"It is a great honor to receive this recognition from the IEEE Board of Directors,” Orso said. “I am thankful to my nominator, references, and endorsers, and I am humbled by being in the company of such great colleagues both within and outside Georgia Tech."

Lee and Orso are two of eight IEEE fellows in SCS. They join Professors Mostafa Ammar, Tom Conte, Ling Liu, Umakishore Ramachandran, Vivek Sarkar, and Ellen Zegura.

The web-based dashboard also lets users compare their predictions to professional models, giving users insight into how to build better forecasts.

The problem with forecasting

Accurate forecasting models are vital to pandemic safety because they can influence public policy and healthcare resource allocation. Yet Covid-19 has been especially hard to predict because its spread is affected by region, season, and other factors.

The idea of combining forecasts comes from forecasting competitions traditionally used to determine the outcome of elections, or the success of businesses, where rewards are given for accurate predictions.

“Competitive forecast mechanisms can be very useful for making predictions in uncertain environments,” College of Computing undergraduate researcher Gayeon Yoo said. “A number of studies have shown the benefits of trusting the ‘wisdom of the crowd.’”

Aggregating the data

The website combines various Covid-19 databases and predictions from the Institute for Health Metrics and Evaluation (IHME), Columbia University, and Georgia Tech for case and death totals. The site determines their accuracy by calculating the mean squared error of the models by prediction time: a week, two weeks, four weeks, and overall. The most accurate forecasts are ranked on the Top Forecasters page.

Users can also provide their own forecasts and learn later how their forecasts performed relative to other users and professional modelers.

“Seeing how your predictions compare to others on different outcomes such as daily cases or daily deaths is what intrigues users,” New York University undergraduate research Rachel Ombok said. “Being able to see the shift in your own predictions and see what your forecasts over time can be useful to researchers, and also helps to engage users who may have private information about the pandemic’s spread in their region.”

The team believes the ranking system encourages users to develop better prediction models by creating a competitive environment around forecast accuracy.

Future iterations of the site will include more data to forecast on hospitalizations and infections, as well segment by country, state, and other regions. The team also plans to make the interface more customizable so users can write their own programs and use statistics or machine learning for automated submissions.

“We want people to see what forecast modeling is like first-hand, and we also want to use our evaluation metrics to quantify the accuracy of user and institution forecasts,” College undergraduate Aniruddha Murali said.

The website was created by College of Computing undergraduates Muraliand Yoo and New York University undergraduate Ombok.

The project is part of a Covid-19 Research Experience for Undergraduates (REU) started by School of Computer Science Assistant Professor Jake Abernethy with School of Computational Science and Engineering Associate Professor Aditya Prakash, University of Michigan Assistant Professor Eric Schwartz, and University of Colorado Assistant Professor Bo Waggoner. The project was funded by the National Science Foundation and Georgia Tech’s Institute for Data Engineering and Science (IDEaS).

Assistant Professor Richard Peng and Professor Santosh Vempala, who holds the Frederick Storey Chair in Computing, discovered that by combining symbolic computing tools with random matrix theory, sparse linear systems can be solved slightly faster than directly invoking matrix multiplication / inversion.

SODA is the premier conference in algorithms and is known to be highly competitive. This year’s conference received 637 submissions and accepted 180 papers. Although up to three papers can be selected, the program committee chose just one paper for the best paper award. 

The history of linear systems

Solving a linear system is one of the most foundational computational problems in science and engineering.

“This problem has been around for a long time, even before computers,” said Vempala. “Much of modern computation runs on linear systems — including recommendation engines like web search, machine learning, and scientific computing simulations.”

Algorithms for solving systems of linear equations are taught as early as grade school, with the most common ones involving manipulating matrices, which in turn represent coefficients on the variables. Yet as their applications become more complex, these methods become less efficient as the matrices become more costly to manipulate.

“The matrix-based algorithms work fine when you have a room full of mechanical calculators trying to solve a system on 50 variables, where the matrix fits on a single large blackboard,” said Peng. “One of the big differences in modern data is this matrix gets really big: instead of 50x50, it’s a billion x billion. Most of the entries start off as 0s, so it’s often not even feasible to write down dense matrices such as the inverse due to storage constraints.”

One example of a large sparse matrix is modern networks. The matrix for a website has every vertex link to another website, creating a gigantic matrix. This matrix is also naturally sparse: it’s mostly made up of 0s. Similar situations also arise in everything from weather simulation to measuring road traffic. As a result, tremendous computing power, including supercomputers,  has been devoted to solving linear systems and problems related to them. For a long time, problems related to solving linear systems routinely consumed the most computational power, until the recent emergence of blockchains.

Since the 1970s, computer scientists have preferred using iterative methods to solve large sparse linear systems due to the difficulty of storing dense matrices corresponding to inverses. Surprisingly little is known about the best possible runtime guarantees of these algorithms on arbitrary systems; analyzing iterative methods is particularly complex because more and more bits need to be maintained with each step to ensure overall accuracy. 

Speedups via batching

Peng and Vempala realized they could solve the systems using fewer steps by considering many partial solutions in parallel. Instead of iterating one vector at a time, they iterate on a block of s vectors. This reduces the number of iterations, and effectively the overhead in bit complexity, from n to n / s.

On the other hand, the interactions between the batches of s vectors across different steps become more complex. Instead of a single value, they become s-by-s blocks. 

More surprisingly, the interactions across steps are also structured. In the Gram matrix, which measures the pair-wise projections between steps, the interactions between steps that add up to the same value are exactly the same. Effectively steps 1 and 7, steps 2 and 6, and steps 3 and 5 all have the same interactions.Turning this idea into an algorithm relies on generalized tools developed from symbolic computing, namely for Hankel matrices, to work in this more general block setting.

Semi-random matrices

Symbolic computing methods are limited because they only work under exact arithmetic. Many previous works studied systems over finite fields, where there are no round-off errors. So existing literature does not address the bit-complexity issue at the crux of the runtime efficiency of iterative methods.

In order to bound the bit complexity of intermediate steps, a major undertaking in this research became lower bounding how close the block Krylov space is to being degenerate, or equivalently, how stable is the computation. A simplified version of this phenomenon arises inpicking n random numbers between 0 and 1: the closest pair should be roughly n^-2 apart. For the block Krylov method, however, this analysis has to be done to the batches of high-dimensional vectors that arise throughout the iterative steps.

The research critically builds upon ideas developed from smoothed analysis: the study of near-degeneracies in problems under small random perturbations. This is a field of study going back two decades, originating with the study of perturbed linear programs.

Wide implications

The new algorithm is a significant contribution to not only linear systems solving, but also the understanding of optimization and random matrix theory. “The  conjecture was that you cannot solve a system faster than matrix multiplication,” said Vempala. “The fact you can go beyond that opens up entire new lines of research.”

The researchers will present the paper, Solving Sparse Linear Systems Faster than Matrix Multiplication, at SODA Jan. 10 through 13.

The School of Computer Science (SCS) Ph.D. student received one of five fellowships for 2021-2022 from a pool of more than 350 applicants. The fellowship includes $50,000 in research funding, a NVIDIA graphics processing unit (GPU), and a guaranteed summer internship.

“The fellowship is a great honor,” Damani said. “I am very excited about the opportunity to go back to NVIDIA, and further my research on GPU compiler optimizations.”

This is not Damani’s first experience with the company. She was an engineer at NVIDIA for five years before starting her Ph.D. at Georgia Tech in 2017. Her work enables programmers to exploit GPU hardware features to improve performance. She specifically works on compiler optimization, which transforms a program to run with improved performance on a specific architecture.

GPUs are used to accelerate deep learning, gaming, visual effects, and high-performance computing applications. They also power 40 percent of the TOP500 supercomputers. As a result, compiler optimizations for GPUs can speed up a wide range of applications.

One thread in her research has been to redesign classical compiler optimizations so that they can be applied to the unique hardware characteristics of GPUs. As an example, Damani redesigned common subexpression elimination, a well-known compiler transformation, so that it can be used to identify duplicate code across diverged threads in GPUs and instead execute them in a converged parallel code region. This optimization results in a reduced execution time.

Another project she has worked on is speculative reconvergence. This programmer-guided compiler optimization improves performance by maximizing parallelism within time-consuming code blocks.

Damani’s research has also been used to accelerate real-time ray tracing, a computer graphics technique for creating realistic lighting in video games, as well as Monte Carlo simulations used in physics, finance, and medicine. She is advised by SCS Chair Vivek Sarkar and is a member of the Habanero Extreme Scale Software Research Lab.

“I was delighted to hear that Sana had been selected for this highly competitive fellowship award,” Sarkar said.  “It is well-deserved recognition of her accomplishments and future potential.”

The fellowship awards partial funding for four Ph.D. students working on novel research topics that fit in with CRNCH’s mission of post-Moore software and hardware designs. Fellows are required to create a poster and paper during their term, and also present at the annual CRNCH Summit in January of each year. This gives them the opportunity prepare for publication and the job market.

CRNCH is a research center that focuses on exploring new computing paradigms after the end of Moore’s law, sometimes called the post-Moore’s era. The center partners with academics and industry to explore full-stack solutions on everything from quantum computing to approximate computation.

Applications for the spring CRNCH fellowship are due Dec. 18, 2020.

Meet the fellows:

Dingtian Zhang
School: School of Interactive Computing
Advisor: Professor Gregory Abowd

Why did you apply for the CRNCH fellowship? 
I am developing computational materials that can weave into the fabric of everyday objects. My work falls under the categories of analog computing, computing based on novel device physics and materials, and optical computing, which is in line with the interest of CRNCH.

What project will you be working on during the fellowship?
We are particularly interested in developing large-scaled sensing systems that can perform light-based sensing on the surfaces of everyday objects to detect implicit and explicit human activities. Such systems need to be self-sustained and easy to maintain, cost effective to scale, conformal to everyday objects, and protective of user privacy. Conventional vision systems based on cameras struggle to keep up with the ubiquitous deployment on these dimensions.

We are developing computational photodetectors that not only sense, but also process the signal in the analog domain to extract mid-level vision features, reducing the inherent complexity and latency from digital signal acquisition and computing. This does not only make the system low-power and scalable, but also prevents capturing unwanted information from images. We adopt emerging organic semiconductor (OSC) devices in fabricating computational photodetectors with lightweight, thin, flexible, and conformal form factors. Computational photodetectors will enable a wide range of large-scale applications such as smart environment, health monitoring, asset tracking, and activity recognition.

Muliang Zhu
School: School of Electrical and Computer Engineering
Advisor: Professor Ali Adibi

Why did you apply for the CRNCH fellowship?
Being part of a larger community that focuses on new frontiers of computing technology is a great benefit for all people like me in the optical computing area. Because of this, I applied for the fellowship to bring the concept of computing using ultracompact photonic devices to CRNCH.

What project will you be working on during the fellowship?
I will be working on nanostructure optical parametric oscillators (OPOs) for nonlinear control of light at the subwavelength scale, aiming at using photonics for neural-network-type computing. The main part of the project I am currently focusing on is the development of nonlinear meta-structure that can provide the optical nonlinearity that is needed for the development of any brain-inspired computing.

Chunxing Yin
School: School of Computational Science and Engineering
Advisor: Professor Rich Vuduc

Why did you apply for the CRNCH fellowship?
My research focuses on neural networks compression using tensorization, which offers a systematic way to trade-off storage, execution time, and accuracy with respect to the capabilities of a given hardware platform. My advisor and I believe that this work fits well within CRNCH and would benefit from feedback from the CRNCH community, so we applied for this fellowship.

What project will you be working on during the fellowship?
We propose to evaluate to what extent convolutional layers and embedding layers of recommender systems can be trained in a reduced form using the techniques of low-rank tensor train decomposition.

Recent studies have shown an alarming growth in the environmental burden from AI, for example the number of parameters in state-of-the-art language models increased to over 175 billion for OpenAI’s GPT-3. To significantly reduce the environmental footprint of AI, we need order-of-magnitude reduction in the infrastructure demand while maintaining or even outperforming state-of-the-art model accuracy. We are exploring a new algorithmic approach, tensor train decomposition, to cope with the large memory requirement of DNNs. The core idea is to replace large weight tensors with a sequence of small tensor decompositions that trades of memory storage with computation. Initially, we will study the compressed networks in the context of heterogeneous CPU-GPU architectures. But we believe that our results will help guide engineering co-design of future hardware-software systems for neural networks.

Ahamad is a professor in the College of Computing and Pindrop co-founder and chief scientist. Along with being a Georgia Tech alumnus, Balasubramaniyan is co-founder and CEO of Pindrop.

The paper, titled PinDr0p: Using Single-Ended Audio Features to Determine Call Provenance, was a team effort supported by Balasubramaniyan, Ahamad, Aamir Poonawalla, Michael T. Hunter, and Patrick Traynor. According to Balasubramaniyan, it has had a significant impact on security research and practice.

“We are so honored to have received this award,” said Balasubramaniyan. “This shows the impact this research has had in a world of instant gratification.”

Published in 2010, the research details the team’s novel approach to identifying and characterizing networks used to make calls in order to create detailed fingerprints for a call’s source. This approach can distinguish between calls made using cellular, internet, or land lines from locations across the world with over 90 percent accuracy, which is the first step in accurately determining the provenance of a call.

PinDr0p was inspired by a trip to India and a call from Balasubramaniyan’s bank to verify a charge to one of his credit accounts.

“I received a call to verify a transaction,” said Balasubramaniyan. “They wanted information to verify my identity — social security number and date of birth — however, I wanted to verify that it was truly the bank calling me.”

The idea took off from there and the team now works with eight of the top 10 national banks and continues to expand its research features­ — now at over 1,300.

"We're building on research that is now established as a Test of Time paper," said Balasubramaniyan. "It's really amazing."

Now as its new chair, Sarkar wants to increase the scale of the school’s research leadership.

“Georgia Tech is a place where we can do big things at scale, make big bets, and be very supportive of our colleagues,” he said of taking on the role.

Finding a future in post-Moore

Sarkar started his career in parallel computing at IBM, but he saw new potential with the end of Moore’s law — the expectation that transistors double annualy in a commodity semiconductor chip.

“The entire computing ecosystem that we have right now in the cloud is predicated on rich system software stacks contributed to by worldwide open-source communities,” he said. “As hardware gets more complicated, a primary concern is that it becomes harder to program in ways that are only accessible to a small number of people with advanced degrees.”

While at Rice University, he created his research group, the Habanero Extreme Scale Research Laboratory, to develop an entire software stack for the new post-Moore world of heterogeneous hardware. Now, at Georgia Tech, the Habanero lab includes five research faculty members and 11 Ph.D. students.

Partnering with faculty in this area was vital for his research into software for future hardware, and what attracted Sarkar to the College.

“One of my big draws for coming to Georgia Tech was having colleagues who are world experts in post-Moore computing,” he said. “Tom Conte’s creation of the Center for Research into Novel Computing Hierarchies (CRNCH) provided a great framework for my research group when I arrived.”

It didn’t take long for Sarkar to participate in multiple collaborative research efforts on campus. Conte invited Sarkar to become the co-director of CRNCH to help broaden the center’s research scope. Sarkar also led a Defense Advanced Research Projects Agency (DARPA) $4.5 million project on software-defined hardware with Conte and Professor Richard Vuduc at Georgia Tech, and other faculty at University of Illinois, University of Michigan, and University of Southern California.

He has other research collaborations under way with fellow SCS faculty: Assistant Professor David Devecsery, Associate Professor Taesoo Kim, Senior Research Scientist Jeffrey Young, and Assistant Professor Qirun Zhang, as well as a the School of Electrical and Computer Engineering Assistant Professor Tushar Krishna.

[Related Content: College of Computing Professors Receive DARPA Contract Award to Improve Software and Hardware Co-optimization]

Serving the community

Yet service is just as important to Sarkar as research. He has always served both the global and local computing community.

“When you are part of a community, it’s great to be able to contribute in every way you can,” he said.

Sarkar is on the Computing Research Association’s Board of Directors and the U.S. Department of Energy’s Advanced Scientific Computing Advisory Committee (ASCAC). He is also an Association of Computing Machinery Fellow and was recently named an IEEE fellow.

Yet as the Stephen P. Fleming Chair in Telecommunications, the service he does within the school and College has been just as important to him. In 2018, he served on the school’s Faculty Recruiting Committee and chaired it in 2019. He also was chair of the School Advisory Committee and a member of dean’s search committee.

Becoming chair of SCS is another form of service for him.

“Service is leadership, and leadership is service — it’s integral to what we do as professors,” he said.

As a leader of SCS, Sarkar wants to bring the same collaborative spirit he’s found throughout the College.

“One of my goals is to really boost our leadership in the core areas that our school contributes to the college that span architecture, databases, networking, programming languages, software engineering, systems, and theory,” he said.

“I’m really keen to figure out how we can integrate research areas where appropriate to put together some really amazing and unique demonstrations of what can be done at Georgia Tech,” he said.  “We’ve hired eight junior faculty members in the last three years who span all these core areas, and have added a new level of energy to the school.”

Watch the interview:

https://youtu.be/_TB_xhl1Grs

Held from Nov. 9 to 10, the University of California, Berkeley workshop is aimed at increasing women in academic careers in computer science, computer engineering, and electrical engineering. Doudali was one of 152 women from 68 institutions who was chosen for her academic excellence, faculty career goals, and equity and inclusion work. The women were mentored by 87 from top-10 institutions.

“Getting selected as a rising star provides me with the endorsement of an elite selection committee that my research contributions and skills are appropriate for me to succeed in academia, boosting my confidence and candidate profile,” Doudali said.

Doudali’s research focuses on building machine intelligent data management for systems with hybrid memory components. The novelty, impact, and practicality of her approach led to a best paper award finalist at HPDC ’19.

[RELATED: SCS Ph.D. Student is at the Forefront of Emerging Memory Systems Research]

“I received valuable advice on how to navigate the academic job search and career, and made wonderful connections with peers and mentors,” Doudali said. “I am grateful for the opportunity and recognition.”

She was also selected to present her thesis at the SC ’20 Doctoral Showcase, has received diversity grants to attend Usenix ATC ’20 and OSDI ’20, and presented her latest publication in MEMSYS ’20.

The Supercomputing Conference (SC), held from Nov. 9 to 19, is the premier event for this rapidly growing field. Mathematicians, engineers, and developers come together to present the most groundbreaking HPC research, and each year, the Georgia Tech name is one of the leading names in those presentations.

Although it is being hosted virtually, this year is no different for SC20 presentations. With several papers, awards, and workshops, Georgia Tech, represented by the School of Computational Science and Engineering (CSE), Partnership for Advanced Computing Environment (PACE), School of Computer Science (SCS), and the Georgia Tech Research Institute, is leading charge in the HPC discussion.

This year, Georgia Tech’s SC20 proceedings includes two notable awards. 

SCS Chair Vivek Sarkar is being recognized with the ACM-IEEE CS Ken Kennedy Award, Nov. 16. Sarkar’s recognition honors his leadership in several areas including foundational technical contributions to programmability and productivity.  

[Related News: ACM/IEEE Recognizes Chair's Research, Service to Computer Science Community]

CSE Ph.D. student Vijay Thakkar and Professor Rich Vuduc are being recognized with a Gordon Bell Prize nomination for their paper, Scalable Knowledge Graph Analytics at 136 PetaFLOPS. Their work, developed with a team from Oak Ridge National Laboratory, is motivated by data mining large-scale bodies of scholarly publications to discover concepts that relate to one another. The announcement of the Gordon Bell Prize winner for 2020 will be held, Nov. 19 from 2 to 2:30 p.m. EST.

An invited talk by SCS and School of Electrical and Computer Engineering Professor Tom Conte is another notable agenda item. Conte’s talk, HPC After Moore’s Law, discusses the post-Moore computing era as seen from the IEEE Rebooting Computing Initiative perspective and presents promising technologies to keep a close eye on.

Excitingly, Georgia Tech students are also participating in a virtual student cluster competition (VSCC) this year. While this event hosts the non-stop 72-hour track of a traditional hackathon, the VSCC separates itself from the rest in that the students prepare up to 6 months in advance for the competition. 

The Georgia Tech team, Team Phoenix, with sponsorship from vendor partner Penguin Computing, will compete with students from around the globe to complete a set of benchmarks and real-world scientific workloads from Nov. 8 to 11. Each team is tasked to design and build virtual clusters in the Microsoft Azure cloud, learn scientific applications, and apply optimization techniques for their chosen cloud configurations. Daily video interviews with the team can be watched here.

While the conference’s papers, presentations, and competitions are underway, this extraordinary year has required extraordinary efforts to transcend the physical conference barriers. With its first ever fully virtual format, SC20’s platform is bringing the HPC discussion into the hands of more than ever before.

To capitalize on this opportunity of enhancing equity in HPC, Georgia Tech is hosting special virtual programming to run parallel to the SC20 agenda. This programming includes a virtual party and poster show that is open to the public – while space lasts – and a 360-degree data center tour. 

The virtual data center tour will be featured on the PACE website and allows users to navigate through Georgia Tech’s premier data center housed at the Coda Building. While walking through the data center, viewers will see the Institute’s two premier HPC resources, the Hive supercomputer and the newly installed Phoenix cluster.

The device, called 6Fit-A-Part, performs wireless localization using ultra-wideband radios to determine if two or more devices are in close proximity. Within seconds of contact, the device emits a red light and a beep to alert the wearer. It can be worn by users or placed on objects in hospitals, schools, grocery stores, warehouses, and similar locations.

Pairing Wireless Networking and Sensing

6Fit-A-Part uses ultra-wideband radios, a relatively new technology gaining popularity in the latest smartphones. Radio waves are ideal for estimating distances because they travel at a fixed speed that can be measured. To determine which device the radio waves sensed the researchers rely on distributed ranging protocol, a wireless networking concept that extends the reach of a signal.

“We have previously explored various ways to measure distance between two wireless devices by exchanging special ranging packets,” said School of Computer Science (SCS) assistant professor Ashutosh Dhekne. “We identified that similar distance measurements can help us raise an alarm if two or more devices come in close proximity to each other.”

In operation, 6Fit-A-Part invites all neighboring devices to exchange wireless ranging packets to determine distance. Messages are limited, so each round of distance measurements isn’t time-consuming. For example, 10 nearby devices can calculate all-to-all distances within a second. The distance is accurate within 10 centimeters if there are no human or building obstructions.

The device’s merging of wireless networking techniques and wireless sensing technology also makes it more precise than other comparable devices. Instead of alerting a user any time devices are in range, 6Fit-A-Part can distinguish if there is a physical barrier between devices, such as a wall, that would prevent close contact.

It can also observe the received signal and its echoes from nearby objects to determine if there was a human blocking the signals. It can then compensate for the distance accordingly. This prevents false negatives when two people are too close to each other, but their bodies obstruct the signals from their devices.

Contact Prevention Better than Contract Tracing

Although many organizations are using contact tracing applications, the focus on social distancing makes 6Fit-A-Part a preventative measure before users are exposed to the virus. Users also don’t need a specific app downloaded or Bluetooth enabled to receive the benefits.

It protects privacy better, according to researchers.

“6Fit-A-Part reminds everyone to maintain physical distancing, which reduces the possibility of infection without constantly tracking a person’s whereabouts,” said SCS Ph.D. student Yifeng Cao. “This allows complete protection of the user’s privacy by not associating a particular person with the device.”

The researchers hope to test 6Fit-A-Part in additional real-world environments to understand how the device influences human behavior. Dhekne and Cao wrote the paper, 6Fit-A-Part: A Protocol for Physical Distancing on a Custom Wearable Device, with SCS Professor Mostafa Ammar. They presented the research at the 28th IEEE International Conference on Network Protocols in October.

By testing the tools in Fulton County, Georgia’s largest county, the Georgia Tech team will be able to solve problems that might be shared by other jurisdictions in the country. The tools will ultimately be available to the general public and election officials nationwide so that people can better understand how public elections are conducted, which increases confidence in their outcome, according to the researchers.

One of the early efforts of the group is to measure and report live wait times to voters at the 250 polling locations in Fulton County on election day, Nov. 3. The website wait.gatech.edu will go live that day and allow users to easily search for a location and view the current estimated wait time. The site also displays wait times recorded throughout the day.

The site is designed to be a utility for voters to help them plan when to vote, and that goal has informed the simple design and usability of the site, according to Ellen Zegura, professor of computer science and a team lead in the research group.

Zegura emphasized that the site does not forecast wait times into the future but rather gives current results based on voter text responses and volunteer observations.

During early voting for the 2020 general election, a student-led team implemented a pilot test at four polling locations in the county. Signs at the polling locations prompted voters to text their wait times after voting. An optional survey let them share more details about equipment quality, Covid-19 concerns, and more. Vlad Kolesnikov, associate professor of computer science who studies cryptography, said the team is mindful of data security and privacy and that no personally identifiable user data is collected or stored.

“By expanding the scope of safe and secure elections from narrow technological problems to addressing physical access, availability, and public health, this project brings a new dimension to the design of modern voting systems,” said Richard DeMillo, principle investigator for the project and chair of Georgia Tech’s new School of Cybersecurity and Privacy.

The election-day effort ties into a larger six-month project by the group that is supported by the Public Interest Technology University Network. It focuses on understanding the quantitative tradeoffs that local election officials are forced to make, and provides tools to help them better manage the voting process. Faculty and students involved in the effort come from the schools of Industrial and Systems Engineering (ISyE), Computer Science, and Cybersecurity and Privacy, as well as the College of Design.

A second project team, led by ISyE Professor Benoit Montreuil, ISyE Director of Professional Practice Dima Nazzal, and ISyE Adjunct Professor and IMT Mines Professor Frederic Benaben, is constructing 3-D maps and AI-based simulations of Covid-safe layouts for a number of Fulton County’s polling places. Their work also involves optimizing voting equipment allocations across all polling locations to minimize wait times, and projections of turn out and waiting times based on historical data and planned equipment provisioning. 

“Georgia Tech’s project is a comprehensive effort that puts much-needed tools and design methods in the hands of public officials that not only improve the process of voting but is a vehicle for communicating to the general public that elections can be both fair and safe in these times of public health crises and social unrest,” said DeMillo.

Michael Best, professor in the Sam Nunn School of International Affairs and School of Interactive Computing, is another team lead and has a long track record of working across the globe to help build sustainable election systems.

“The goal is very straightforward. We hope to enhance safety, security, and efficiency at polling locations, which of course should contribute to better voter turnout and trust," said Best.

Volunteers can help the research group in providing live wait times at polling locations to Fulton County voters by observing voters on Nov. 3 and texting wait times, or by helping set up signs Nov. 1 so voters can directly text wait times.

Qureshi joined the School of Computer Science in July to bring his expertise in this area.

The Quantum Connection

Quantum wasn’t always Qureshi’s research area. He received his Ph.D. from University of Texas at Austin, where he studied cache management. He then worked at IBM on emerging memory technology. When he started at the School of Electrical and Computer Engineering in 2011, Qureshi’s areas were memory systems, computer architecture, and hardware security.

It wasn’t until 2016 when a student came back from a Microsoft internship with a new interest in memory system for quantum computers that Qureshi started working in the area. Until 2016, research in quantum computing had mostly a theoretical bent, and Qureshi was skeptical of the practical implementation and use. That all changed once IBM made the 5-qubit machine available via cloud for public use and users could run programs on real quantum computers.  Since then, companies such as IBM and Google have demonstrated machines with 50+ qubits, and the progress in machine capacity is expected to accelerate in the coming years.

“My background is computer systems, not quantum physics or quantum information,” he said. “Before 2016, there was no obvious bridge to do system-level research in quantum computing”

“Quantum computers became really exciting for system research with the advent of the cloud-based IBM quantum machines,” Qureshi said.

Moving from Engineering to Computing

As Qureshi’s research shifted more into quantum, it became clear he could make more headway in the field if he joined the College of Computing, where it would be easier to work with researchers in programming languages and software.

“The chance of collaboration, the chance of learning and doing impactful research, and teaching quantum computing courses is higher in College,” he said.

Qureshi’s goal is to improve software to make quantum computation less susceptible to errors. Some of his recent work uses diversity to reduce the errors in quantum computation. Qureshi also led a collaboration with Oak Ridge National Labs (ORNL) that allows Georgia Tech employees to have access to the state-of-the-art quantum computers from IBM.

Qureshi wants to make quantum accessible to students.

“The content on quantum computing can be intimidating, and sometimes it may feel like you need a Ph.D. in physics and another Ph.D. in mathematics to make sense of the material” said Qureshi. “However, if you are interested in doing research at the system and software level, then the problems are similar to typical systems and abstractions can go a long way in making progress.”

With this in mind, Qureshi is teaching an Introduction to Quantum Computing graduate class this fall and hopes to bring quantum to the Online Master of Science Computer Science (OMSCS) program by 2021.

“The exciting aspect of doing systems research in quantum computing is that the field is only getting defined right now, and you have more of a chance of making a bigger contribution in a field of when you go in early,” he said.

The award honors research that had the most impact out of the papers published that year. Orso and Halfond, who is now an associate professor at the University of Southern California, won for their paper, AMNESIA: Analysis and Monitoring for NEutralizing SQL-injection Attacks.

Amnesia is a fully automated technique for detecting and preventing one of the most catastrophic types of web application attacks.

The research

SQL injection attacks (SQLIAs) inject malicious code into databases to expose information. This can lead to private information being leaked or even entire databases being corrupted. SQLIAs are one of the most prominent attack types, and at the time of this research, were considered the number-one threat for web applications.

Before Orso and Halfond introduced Amnesia, developers had to manually incorporate specific checks into their applications. This process was both time-consuming and prone to error.

Amnesia was the first fully automated techniques for detecting and preventing SQLIAs that was widely applicable and successful.

“Our approach was based on the intuition that developers implicitly provide, in the web application code, a policy on what kind of database requests are allowed,” Orso said.

With this in mind, Amnesia’s approach did three things:

1. Extracted a policy from the code using static analysis
2. Checked database requests against this policy
3. Stopped requests that violated the policy, as they were likely SQLIAs

The impact

The paper made ripples in the program analysis community.

“Our paper was one of the first papers that successfully applied program analysis techniques to the problem of SQLIAs,” Orso said.

As a result, other research groups built on that work and its underlying idea. To date, the original paper has been cited over 700 times.

It also jump-started Orso’s career. The concept became the basis for a project sponsored by the Department of Homeland Security, Preventing SQL Code Injection by Combining Static and Runtime Analysis, in collaboration with Professor Wenke Lee.

Orso and Halfond continued to advance the SQLIAs detection and prevention area in both their careers. The work also motivated Orso’s research group to develop general testing and analysis techniques for web applications — work that ultimately became Halfond’s Ph.D. dissertation.

“Receiving this prestigious award from the research community for a paper already so close to my heart is a humbling, exciting, and incredibly rewarding experience that goes beyond my wildest expectations,” Orso said.

The Kennedy Award was established in 2009 to recognize substantial contributions to programmability and productivity in computing and significant community service or mentoring contributions.

"I am delighted to receive this year’s Kennedy Award," Sarkar said. "I have always believed that research and mentoring go hand in hand, and it is an honor to be recognized for my work in these areas.  The award is also meaningful because of the influence that Ken had on my career."

[RELATED: Sarkar Named as School of Computer Science Chair]

Sarkar began his career in IBM Research in 1987 after obtaining his Ph.D. from Stanford University supervised by John Hennessy. His research projects at IBM include the PTRAN automatic parallelization system led by Fran Allen, the ASTI optimizer for IBM’s XL Fortran product compilers, the open-source Jikes Research Virtual Machine for the Java language, and the X10 programming language developed in the DARPA HPCS program. He was a member of the IBM Academy of Technology from 1995 to 2007.

“Vivek is not just a leader in research, but also a generous teacher and mentor, which is to say he is a role model for our community,” said Charles Isbell, dean of the College of Computing and John P. Imlay, Jr. chair.

Since moving to academia, Sarkar has mentored more than 30 Ph.D. students and postdoctoral researchers in the Habanero Extreme Scale Software Research Laboratory, first at Rice University and then at Georgia Tech. While at Rice, Sarkar was the E.D. Butcher Chair in Engineering, served as Chair of the Department of Computer Science.

The Kennedy Award recognizes Sarkar’s leadership in several areas including foundational technical contributions to programmability and productivity. Sarkar has developed innovative programming-model, compiler, and runtime technologies for parallel computing that have influenced other researchers, as well as industry products and standards. Researchers in his lab have developed the Habanero-C/C++ and Habanero-Java programming systems for parallel, heterogeneous, and distributed platforms. These systems introduced new constructs for homogeneous and heterogeneous intra-node parallelism, as well as their integration with scalable inter-node communication libraries.

Sarkar has led open-source software projects that have had a significant impact on the research community: he has created new pedagogic materials to make parallel programming more accessible to undergraduate students and the Coursera learner community and has mentored junior colleagues at IBM and several Ph.D. students after moving to academia.

Sarkar has also demonstrated leadership in community service by serving as program chair and general chair for major conferences in his research area, serving on U.S. Department of Energy’s Advanced Scientific Computing Advisory Committee (ASCAC) advisory committee since 2009, and on the Computing Research Association (CRA) Board of Directors since 2015.

The Kennedy Award carries a U.S. $5,000 honorarium endowed by the IEEE CS and the ACM, which Sarkar plans to donate to Georgia Tech’s Constellations Center for Equity in Computing.

“Vivek is not just a leader in research, but also a generous teacher and mentor, which is to say he is a role model for our community,” said Charles Isbell, dean of the College of Computing and John P. Imlay, Jr. chair. “We are lucky to have him as a leader.”

Currently, people rely on blacklist apps to stop spam calls, but they are only up to 60 percent effective, according to School of Computer Science Ph.D. student Sharbani Pandit. This is because of the prominence of neighbor spoofing, or numbers similar to person’s own.

“The caller number will look very similar to your own number, so you’re more likely to pick up, and each time they call, they use a new number, so it doesn’t show on the blacklist,” said Pandit.

So Pandit and her team tried a more direct approach, a virtual assistant (VA). Like a smart home device or a secretary, the VA determines whether calls are from people before it passes the call off to the user.

How It Works

Whether a call is spam or not can be determined by a simple question: does the caller know the name of the person they’re calling?

“The idea is someone who is a wanted caller would know the full name of the person they’re trying to reach,” Pandit said.

If the caller can confirm the name, the VA would forward the call and a transcript of how the conversation went to the user. The entire interaction takes 10 seconds.

For callers who may not know the name but are not a bot, VA engages in a conversation and interrupts the caller as they are speaking to determine if the caller stops talking. This process takes a maximum of 30 seconds.

“It’s very natural in human conversation that you would stop to listen to what the other party is saying on the call, but a bot wouldn’t do that,” Pandit said.

For calls thought to be spam, the notification would be sent to the user with a label for each caller to indicate whether it was a human or a robot caller.

How Many Calls It Stops

The researchers did a user study with 21 people who determined they were comfortable talking to the VA. They also tested the VA on a database 8,000 robocall recordings, in which 97.8 percent of the robocalls were correctly labeled as such.

Pandit presented the research in Fighting Voice Spam with a Virtual Assistant Prototype. She co-wrote the paper with University of Georgia’s Jienan Liu and Associate Professor Roberto Perdisci, and SCS Professor Mustaque Ahamad.  The work has already garnered media attention from New Scientist and Digital Trends.

The researchers are still making a few adjustments before they release the Android app. Robocalls are constantly evolving and getting more sophisticated, which introduces new challenges to researchers. Some robocalls even have lists of phone numbers and names, so the researchers are adding more questions in the screening conversation that only a human could answer.

They believe that anything that can slow down the efficacy of robocalls will be a benefit.

“The goal of robocalls is to make as many calls as possible to target victims and make a profit from it, but as security researchers, if we can add hurdles in their step, then sometimes the cost isn’t worth it for the attacker,” said Pandit.

Although all virtual this year, it’s still one of largest gatherings of women in computing with more than 30,000 people from 115 countries representing academia and industry.

Thanks to scholarships from the College, 63 undergraduate students, 32 master’s students, six Online Master’s of Science in Computer Science (OMSCS) students, and 12 Ph.D. students are able to attend.

They have the opportunity to watch more than 200 panels and keynotes. Some highlights from Georgia Tech include a fireside chat with Joy Buolamwini, an alumna and founder of the Algorithmic Justice League, on Decoding Bias on Oct. 3.

The Constellations Center for Equity in Computing’s Director of Educational Innovation and Leadership Lien Diaz joins the panel Seeing Beyond Yourself: Effective Allyship, Advocacy, and Activism for Women in Computing on Sept. 29.

“I am particularly interested in the wide array of topics that GHC speakers will be addressing from tech careers to applications of machine learning and artificial intelligence,” said OMSCS student Michelle Adea.

The conference is just as much about networking as learning. As a silver-level sponsor, the College will connect with prospective students.

Some students are excited to meet other women in computing.

“I’m looking forward to engaging with other like-minded women in different career positions and levels of education and making connections,” said undergraduate Rashmi Athavale.

“This is the first time in the history of VLDB that we have had such a diverse and dominating presence at the conference,” said SCS Professor Shamkant Navathe. “We hope to continue this tradition of a strong presence at VLDB in the coming years."

These publications encompass many of SCS’s top research areas, including databases, systems, networking, security, and architecture.

“We have papers spanning a wide array of topics including graph systems, cloud computing, video analytics, similarity search, and database fuzzing,” said SCS Assistant Professor Joy Arulraj.

[Related Content: New Toolchain Automatically Finds Database Management System Bugs]

Assistant Professor Xu Chu was recognized as a distinguished reviewer this year.

SCS presented the following papers:

ODIN: Automated Drift Detection and Recovery in Video Analytics
Authors: Abhijit Suprem (Georgia Tech)*; Joy Arulraj (Georgia Tech); Calton Pu (Georgia Tech); Joao E Ferreira (University of Sao Paulo)

iDEC: Indexable Distance Estimating Codes for Approximate Nearest Neighbor Search
Authors: Long Gong (Georgia Tech)*; Huayi Wang (Georgia Tech); Mitsunori Ogihara (University of Miami); Jun Xu (Georgia Tech)

APOLLO: Automatic Detection and Diagnosis of Performance Regressions in Database Systems
Authors: Jinho Jung (Georgia Tech)*; Hong Hu (Georgia Tech); Joy Arulraj (Georgia Tech); Taesoo Kim (Georgia Tech); Woonhak Kang (eBay Inc.)

Traversing Large Graphs on GPUs with Unified Memory
Authors: Prasun Gera (Georgia Tech)*; Hyojong Kim (Geroga Tech); Piyush Sao (Oak Ridge National Laboratory); Hyesoon Kim (Georgia Tech); David Bader (New Jersey Institute of Technology)

A Drop-in Middleware for Serializable DB Clustering across Geo-distributed Sites
Authors: Enrique Saurez (Georgia Tech)*; Bharath Balasubramanian (ATT Labs ); Richard Schlichting (United States Naval Academy); Brendan Tschaen (ATT Labs); Shankaranarayanan Puzhavakath Narayanan (ATT Labs); Zhe Huang (ATT Labs); Umakishore Ramachandran (Georgia Tech)

Cloudburst: Stateful Functions-as-a-Service
Authors: Vikram Sreekanti (UC Berkeley)*; Chenggang Wu (UC Berkeley); Charles Lin (UC Berkeley); Johann Schleier-Smith (UC Berkeley); Joseph Gonzalez (UC Berkeley); Joseph M Hellerstein (UC Berkeley); Alexey Tumanov (Georgia Tech)

A  new open-source tool from Georgia Tech, however, is making research more accessible for everyone.

Vortex is a reconfigurable GPGPU accelerator that uses an extended open-source reduced instruction set computer (RISC), an open architecture. By combining a GPGPU with RISC-V, the hardware can use the open programming framework OpenCL.

“Before RISC-V, building full-stack GPU hardware would have been an insurmountable task that only large companies would have had the resources to do,” said Blaise Tine, a School of Computer Science (SCS) Ph.D. student who leads the Vortex GPU. “One of the main design goals of this project is to support open-source tools across the board; software, compilers, simulators, synthesis, etc., to really make this platform as accessible as possible.”

Hardware roadblocks

Since the number of computer chip transistors is no longer expected to double every year as Moore’s law held up until the past few years, the semiconductor industry has been pushing specialized hardware to make up for performance gaps.

Now heterogeneous systems-on-chip provide powerful performance, but a lot of GPGPU research is restricted to simulation unless researchers can work with large companies on the actual hardware.

“Hardware design is extremely difficult and time-consuming, and before the emergence of RISC-V, there were no standardized ISA or software tools that guaranteed long-term support of open-source contributions like this,” said Tech alumnus Fares Elsabbagh.

Open-source solutions

Yet with the advent of open-source ISAs like RISC-V, and the programming framework OpenCL, the researchers saw an opportunity.

To support OpenCL on Vortex, the researchers used the open-source compiler POCL and extended it to support RISC-V and Vortex ISA. Being able to use OpenCL was a major breakthrough for Vortex because it allowed the processor to execute real applications. This has practical applications, such as running existing machine learning programs without having to rewrite the program.

“Vortex is built on industry-standard open-source tools and protocols, allowing researchers and companies to easily modify the design to their needs, and be able to easily evaluate it against other designs,” Elsabbagh said.

Vortex is already being used in student projects, and the researchers have plans to extend its capabilities. They also plan to expand the platform to cover all aspects of hardware acceleration, graphics, artificial intelligence, vision, and audio.

The researchers presented Vortex in the poster, Vortex: An Open Source Reconfigurable RISC-V GPGPU Accelerator for Architecture Research, at Hot Chips: A Symposium on High Performance Chips from Aug. 16-18. Tine and Elsabaggh wrote the paper with Tech studnets Apurve Chawda, Will Gulian, Yaotian Feng, Da Eun Shim, Priyadarshini Roshan, Ethan Lyons, Lingjun Zhu, School of Electrical and Computer Engineering Professor Sung Kyu Lim, and SCS Professor Hyesoon Kim.

APOLLO automates the generation of regression-triggering queries, simplifies the bug reporting process for users, and enables developers to quickly pinpoint the root cause of performance regressions.

The researchers discovered 10 previously unknown and unique performance regressions, reduced query size by 4.2 times, and identified branches related to the root cause.

DBMS problems

The complexity of DBMS increases their potential for error. An upgrade on a DBMS can unexpectedly slow down certain queries, a problem known as a performance regression bug.

“A critical regression can reduce performance by orders of magnitude, in many cases converting an interactive query to an overnight execution,” said SCS Ph.D. student Jinho Jung.

To improve this issue, the researchers used the toolchain approach, a pipeline of distinct software development tools that are linked together by specific stages.

The team’s new toolchain has three components:

SQLFuzz

SQLFuzz generates structured query language (SQL), the language databases communicate with, to find performance regressions. It works by bombarding a system with many randomly generated inputs to trigger bugs, a technique known as fuzzing.

“During the fuzzing test, we noticed that validating performance regressions is challenging because the ground truth of the regression is unclear and may be heavily affected by the 

execution environment and lead to a lot of false-positive bugs,” Jung said.

To counter this, the researchers applied validation checks to reduce false positives.

SQLMin

SQLMin minimizes the regression-triggering query, so performance isn’t compromised by trying to determine the essence of a regression-causing statement. The researchers achieve this by using both bottom-up and top-down approaches.

Bottom-up strategy extracts one sub-query from the database and monitors whether there is still a regression problem. If there is one, SQLMin keeps the sub-query for further analysis. The top-down strategy removes as many expressions as possible.

“This takes out as many elements of the statement as possible while ensuring that the reduced query still triggers the problem,” Jung says.

SQLDebug

Once a regression report is filed, developers must diagnose its root cause. To simplify the diagnosis process, the researchers use two techniques to automatically identify the root cause.

First, they use the mathematical approximation method of bisecting to find the historical commit, or first code update, that the developer pushed to the code repository. Second, they leverage statistical debugging to determine if performance decreased because of suspicious source lines within the commit.

The researchers introduced Apollo at the Very Large Data Bases conference from Aug. 31 to Sept. 4. Jung wrote the paper, APOLLO: Automatic Detection and Diagnosis of Performance Regressions in Database Systems, with SCS Ph.D. postdoctoral student Hong Hu, Arulraj, and Associate Professor Taesoo Kim, and eBay’s Woonhak Kang.

As the cloud continues to bring easy data storage to billions, operators are pushing their networks to peak efficiency to reduce costs. Datacenters are designed with the expectation that there will be more internal data than external, but as more people join the cloud, this paradigm is no longer true, according to the researchers.

Cloud conundrum

As soon as the cloud became a prominent data storage method, cloud operators started leasing compute, storage, and network resources to their clients. The rise in demand for these resources means that networks are now more overloaded than ever.

Cloud providers’ solution to the influx of computers on the network was to allocate more  resources than necessary. Yet this isn’t cost effective, according to School of Computer Science (SCS) alumnus Ahmed Saeed.

The cloud is made of interconnected datacenters. One machine in a datacenter can send traffic to other machines in the same datacenter and others outside of it. This is further complicated by data transmission speed, or latency, which is very fast inside the datacenter. However, the latency between datacenters is much longer because data must travel across the country through the WAN.

When both types of traffic compete, the faster reacting type of traffic shoulders the burden of reacting to any changes in network conditions, harming overall performance. This problem can be ignored when the demand of the slow traffic is low, but as the demand increases, the problem becomes more prominent.

“When the bandwidth demand of datacenter-to-datacenter traffic is small, it is not very costly to dedicate that bandwidth within the datacenter network,” Saeed says. “However, as that demand increases, dedicating resources can be wasteful.”

If the network is bottlenecked, the datacenter traffic will immediately sense the problem and compensate for it. However, WAN traffic will take much longer. In effect, the entire network becomes 2.5 times less inefficient before all parts of the network are aware of it.

“Information has to travel from one data center through the WAN and back,” said Saeed. “It’s problematic.”

Congestion control

When demand for network bandwidth is larger than available capacity, the part of the network responsible for handling traffic is called congestion control. This is a fundamental concept in network research and how the internet functions. Yet this efficiency issue between datacenter traffic and WAN traffic has led to a new type of congestion, and the solution has to be time focused, according to Saeed.

“It’s not a matter of how good your network signal is or how smart your algorithm is, but how fast you react,” he said.

With this in mind, the researchers worked to the solve the problem in the datacenter, not through the end points. Annulus works by sending a message from the congested part of the network to the traffic source to let it know how to react, essentially cutting out the middleman.

Annulus has two control loops. One uses existing congestion control algorithms for bottlenecks at just one source, either the WAN or the datacenter. The other loop focuses on bottlenecks between the WAN and the datacenter and tries to solve the problem at the traffic source.

When tested, Annulus improves datacenter traffic by 56x when the bottleneck is the WAN. It also reduces datacenter flow issues by up to 3.5x.

Saeed presented the research in the paper, Annulus: A Dual Congestion Control Loop for Datacenter and WAN Traffic Aggregates, at the ACM Special Interest Group on Data Communication (SIGCOMM) from Aug. 10 to 14. He co-wrote the paper with SCS’s Professors Mostafa Ammar and Ellen Zegura, and Varun Gupta, Prateesh Goyal, Milad Sharif, Rong Pan, Keon Jang, Mohammad Alizadeh, Abdul Kabbani, and  Amin Vahdat.

Unlike other heap exploitation techniques that require considerable effort from the researcher, ArcHeap can autonomously explore the system.

“Many heap exploitation techniques have been discovered by researchers; however, this task always relies on manual efforts,” said School of Computer Science (SCS) Ph.D. student Insu Yun. “We wanted to automate this process.”

[RELATED CONTENT: Team IDs Real-world Vulnerabilities In Popular Browser During Premier Hackathon]

Heap exploitation techniques

Heap is dynamically allocated memory, or memory that’s size is determined during program execution. Heap allocators manage it efficiently, yet they are also very susceptible to attack.

Exploitation techniques abuse underlying heap allocator mechanisms to exploit vulnerabilities. Popular systems software is plagued by heap-related vulnerabilities. Microsoft said heap vulnerabilities led to more than half of their security problems in 2017. Heap vulnerability  attacks have also been seen in popular software such as WhatsApp, VMware, and Eximail in 2019.

Since each exploit is specific to the allocator, previous research was specialized and manually done. This created a barrier to understanding security issues with various heap allocators and led to even more attacks.  

How ArcHeap works

ArcHeap introduces operations and attack capabilities to see if these trigger an exploitation. The approach is comparable to fuzzing, an automatic software testing technique that inputs random data to expose vulnerabilities.

“We found that fuzzing is very useful for finding software vulnerabilities, so we extended this idea into discovering heap exploitation techniques,” Yun said. “However, classical fuzzing cannot be naively applied to this new problem, so ArcHeap employs several new ideas.”

The researchers determined heap allocators share common design components that allowed them to abstract enough so that the tool can be applied to any allocator. ArcHeap also synthesizes its finding as it moves through the allocator to reduce redundancies.

ArcHeap’s findings

Researchers evaluated ArcHeap on 11 allocators and found five new exploitation techniques in Linux's default allocator, ptmalloc2. Despite decades of research in this area, ArcHeap successfully discovered heap exploitation techniques in ptmalloc2.

“Our results show that their manual security analysis was insufficient to cover a large space of heap exploitation techniques,” Yun said. “As a result of this insufficient manual testing, these allocators were actually not secure as their claims.”

They also found vulnerabilities in seven of the 10 other popular allocators.

Although right now ArcHeap can determine if a heap allocator is vulnerable or not, the researchers hope to put a quantitative value on that security in the future.

They presented the research in the paper, Automatic Techniques to Systematically Discover New Heap Exploitation Primitives, at Usenix’s 29^th Security Symposium from August 12 to 14. Yun co-wrote the paper with SCS Associate Professor Taesoo Kim and Facebook’s Dhaval Kapil.

“Humans are biased themselves, so how can you resolve an issue of bias with a human?” School of Computer Science Ph.D. alumna Samira Samadi said. “It might even make it worse.”

The limits of facial recognition

Facial recognition software supposedly automates building security. The software takes photos as people enter a building, which it  then cross-references with an employee database. If the software finds a match, a person can enter the building.

Despite many advances in image recognition and artificial intelligence, systems are often more accurate for men with lighter skin tones and less for women with darker skin tones. Companies have proposed adding a human evaluator to compensate for the software’s limitations.

Yet Samadi, who researches algorithmic fairness, immediately recognized the potential for more bias. She wanted to know whether adding a human evaluator to the process increases fairness or bias.

Experimental design

Yet designing such a human/user study is challenging as Samadi and colleagues at Microsoft Research realized. Working with actual security guards or receptionists would be ideal, but was not feasible in practice.

Samadi turned to recruiting people through Mechanical Turk as she had done in the past. These users would offer her volume, but they were not trained in recognizing faces. First, she studied how to compare faces. Then she learned how to teach Mechanical Turk users about facial recognition systems, how to make decisions about the accuracy of the system, and how to be confident in that decision.

After research, Samadi developed a user study and did some trials with friends to ensure the study was clear and understandable. Then she ran the study on 300 users on Mechanical Turk.

Each user was trained on how to distinguish faces and evaluate the software. Next, the user saw two images and how they were scored by the software. Samadi expected the human evaluator would show bias between two lighter versus dark-skinned people, but the results were much different.

Future studies

“We really tried to imitate a real world scenario, but that actually made it more complicated for the users,” Samadi said.

The researchers were unsure whether the problem with the study  was because users didn’t understand the study or biased behavior, but they ultimately decided not to publish the research. However, Samadi did publish a position paper, A Human in the Loop is Not Enough: The Need for Human-Subject Experiments in Facial Recognition, with Microsoft Research’s Farough Poursabzi-Sangdeh, Jennifer Wortman Vaughan, and Hanna Wallach. Samadi presented the work at the Conference on Human Factors in Computer Systems (CHI) in April.

The paper argued about both the necessity and issue with studies like these. There are four main challenges about both the efficacy and generalizability with a human-subject study like the one they conducted:

-Datasets: Finding an appropriate dataset is difficult for a number of factors: Sourcing images ethically is challenging because past research has relied on celebrity or politician images who are easily recognizable and thus bias the study. Many datasets are also already biased and contain more lighter-skinned faces than darker. Also, many datasets are higher quality than what would be found in camera footage and not an effective real world comparison.

-Participants: Many available participants for studies like these are students or Mechanical Turk workers who are inexperienced in facial recognition.

-Context: Recognizing faces in an experiment is not comparable to on the job duties when an unfamiliar person may be a threat.

-User Interface: Companies do not release their user interfaces for facial recognition software, leaving it up to researchers to design something that may not reflect what is used in real world software.

“If someone wants to attack this problem in the future, they should know the challenges they have ahead of them,” Samadi said.

Barua created the control circuit that enables the ventilator to be run remotely.

When lockdown started in early March, Barua’s friend Prasoon Suchandra, a Ph.D. student in mechanical engineering, reached out for volunteers to develop a control circuit for their emergency low-cost ventilator project led by Devesh Ranjan, a professor in the George W. Woodruff School of Mechanical Engineering.

“I had some experience in developing such circuits and programming the controllers, so I decided to help them,” said Barua, who normally works on program analysis and compiler optimizations for SCS Professor and incoming Chair Vivek Sarkar’s Habanero Extreme Scale Software Research Lab.

Power of the control circuit

Since the Covid-19 pandemic started, people across the world have been building makeshift ventilators, but most of them have no or minimal controls and work only on sedated patients.

“Without the control circuit, a medical professional has to manually monitor and control all such parameters for patient safety,” Barua said.

A control circuit makes it possible to control and monitor all the different vitals that doctors need for any ventilator to be useful: respiration rate, tidal volume — a measure of how much air is inhaled in a breath — inspiration and expiration ratio, and pressure on lungs.

Yet the control circuit also empowers the patient. Barua added sensors and a RaspberryPi controller that reads the sensors, performs some numerical computations, and updates a user interface continuously in a loop, eliminating the need for a medical professional to be present.

A patient can support the volume of air delivered, time their breathing, and cough without fighting the ventilator. In effect, the ventilator reduces the involvement of a medical professional.

Future of the ventilator

Making a functional ventilator was one of the goals of the project.

“Our entire team has spent a lot of effort on developing a design that can be really useful in the current situation,” Barua said. “We have interacted with different doctors and hospitals and taken their feedback into account.”

The team has also ensured that it can be manufactured cheaply with easily available raw materials almost anywhere in the world. They hope to openly release the design, so that anyone can mass manufacture it. They are also exploring partnering with a few potential companies around the world to manufacture the ventilator. Whoever manufactures it will get the required certifications and have legal liability for the device.

Research to reality

Although the work doesn’t overlap much with his typical research, Barua applied some approaches he uses in his work on optimizing the memory access patterns on graphics processing units (GPUs). For the ventilator project, he selected, updated, and processed appropriate data structures to optimize the program. He also applied simple parallelization methods like multiprocessing to achieve the right sampling rate for the RaspberryPi.

“It was interesting to see the ventilator being built from scratch in the workshop, working with oscilloscopes to debug our circuit in the electronics lab, and developing programs to control actual mechanical systems,” he said.

For Barua, the main goal was having a positive impact during the pandemic and collaborate on a real-world problem.

“This project was also unique since we were remotely collaborating with a very big team from different departments including the Georgia Tech Research Institute. It was a great learning experience.”

For the past year, the School of Computer Science has been led by interim chair and Regents’ Professor Mostafa Ammar.

“I’d like to thank Mostafa for stepping into leadership when his school and his college needed him,” Isbell said. “He has shown the kind of commitment to community that we all strive to emulate. I truly appreciate his thoughtful efforts during the past year.”
 
Sarkar will start as chair on August 16. Please join Dean Isbell in congratulating him and the School of Computer Science on this new chapter.

A team of Georgia Tech researchers has created a system that allows users to more effectively label a training dataset with higher accuracy than current methods.

“We are looking at the problem from a data management perspective,” said School of Computer Science (SCS) Assistant Professor Xu Chu. “In contrast to a lot of ML research that tries to tackle the lack of sufficient training data from an ML algorithm design perspective, we aim at building a system that helps users effectively label a dataset.”

The system, called GOGGLES, labels datasets using affinity coding, a paradigm that allows ML engineers to use various affinity functions that input two unlabeled examples and output a real-valued score.

“You can think of affinity as similarity,” said Chu. “The core premise of the work is that two examples share the same label if they are similar according to some affinity functions (or similarity functions).”

The benefits of affinity coding

GOGGLES uses a set of affinity functions that can capture various affinities found in the image. Next, using a new unlabeled dataset and these affinity functions, GOGGLES constructs an affinity matrix, from which it can assign classes to unlabeled images. This doesn’t require any metadata or developer intervention like previous .

For each new dataset, users can potentially reuse many of the existing affinity functions already in the library, making GOGGLES a domain-agnostic labeling system. Users and developers can always add more affinity functions to increase the labeling power of GOGGLES.

On five common image classifying tasks, GOGGLES reaches up to 98 percent accuracy without requiring extensive developer effort. It also outperforms other well-known data programming systems by up to 21 percent.

Chu co-wrote the paper, GOGGLES: Automatic Image Labeling with Affinity Coding,  with Ph.D. students Nilaksh Das and Renzhi Wu, master’s alumni Sanya Chaba and Sakshi Gandhi, and School of Computational Science and Engineering Professor Polo Chau. They presented it at Association for Computing Machinery's Special Interest Group on Management of Data (SIGMOD) and Symposium on Principles of Database Systems (PODS) held virtually from June 14 to 19.  

In 2010, a group of researchers at Georgia Tech, the CDC, and Ministries of Health in several African countries launched an open-source laboratory management system as part of the College of Computing’s Computing-for-Good (C4G) initiative. Designed to be ultra-configurable to meet variable needs of labs across developing countries with minimal training for staff, it quickly grew to become one of C4G’s biggest success stories.

More than 10 nations in sub-Saharan Africa adopted the program, called the Basic Laboratory Information System (BLIS), giving areas with little or poor internet connectivity an easy-to-use system for many who had minimal computing experience. These countries, which had over 1 million patients at the time, were using paper-based systems to manage information on disease spread, local illnesses, and much more. As information and communications technologies have expanded in the area, however, many labs gained a standardized reports system that could track prevalence rates of infections, slowing their spread.

But a lot can change in just 10 years. What was once designed for personal computing interfaces is now desired for a wide range of new platforms. Although laptops are still the device of choice for the majority of nurses – 79.6 percent reported in a study of a Nigerian hospital -- smartphones and tablets have seen a steady increase. The coming years will include many more innovations that render even those obsolete.

As users in the global south aspire to embrace mobile computing in clinical settings, a flexible interface, adaptable to everchanging applications, is needed.

Enter: Jung Wook Park and Aditi Shah, a Ph.D. student in Georgia Tech’s School of Interactive Computing (IC) and former master’s student in the School of Computer Science (SCS), respectively. Along with SCS Professor Santosh Vempala and IC Principal Research Scientist Rosa Arriaga, Park and Shah published research updating the current interface of C4G BLIS

Their updates focused on a handful of key areas, primarily mobile support. A responsive user interface framework supporting various screen sizes and resolutions was developed and evaluated by real users at hospitals in Africa currently using BLIS.

They compared user experience with the current interface on desktops and smartphones with a proposed interface on both and found that there was a significant improvement on both the desktop and smartphone.

“When you bring in a new system, they may feel uncomfortable with it,” Park said. “If we didn’t do a great job, you might get the same score or lower at the beginning. Over time, we saw improvements of 32 and 34 percent on desktops and smartphones.”

Shah, now at Microsoft, offered plenty of help in the development of the system, and her experience with a visual impairment allowed her to provide perspective on accessibility, as well.

The implications of this research extend far beyond ease of use for nurses, however. Park identified a growing problem across the globe in health care: communication. As the current pandemic can illustrate, viruses and diseases can spread quickly across many different populations. It isn’t sufficient to have just local data to mount an appropriate response; teams around the world must be able to rapidly share information.

A system like C4G BLIS, with its improved user interface that can be used across multiple platforms depending on the local needs of various communities, can help that communication.

“If you notice something locally and maybe other areas of the country or continent notice something, how do you know if it is a pandemic?” Park posed. “You need to be able to share that information to manage the spread. By turning these local systems into a standardized cloud-based system, we can improve communication.”

Already, Vempala said, he has heard reports from many labs that have adapted the flexible system to keep track of COVID-19 data in their communities.

The paper is titled Redesigning a Basic Laboratory Information System for the Global South, and was presented at the International Telecommunication Union Kaleidoscope conference, earning a Best Paper award.

Boldyreva, who is a professor and associate chair in the School of Computer Science, wrote the single-author winning paper, Threshold Signatures, Multisignatures and Blind Signatures Based on the Gap-Diffie-Hellman-Group Signature Scheme, in 2003 as a Ph.D. student at the University of California San Diego.

“My advisor had to convince me to write the paper because I thought the results were too simple to deserve a publication,” Boldyreva said. “Of course, back then I would have never believed that the paper would do so well.”

Her new multi-user digital signature schemes were simpler and more efficient than existing schemes at the time. One of the most useful was a multisignature that allows a number of users to jointly digitally sign the same message while keeping the final signature as short and computationally efficient as a single signature. In effect, the length of the multisignature doesn’t grow as more users sign.

Secure networking protocols as well as blockchains and cryptocurrencies commonly rely on efficient multi-user signatures. Yet, at the same time, the schemes Boldyreva created are still simple enough that introductory cryptography courses teach them.

Boldyreva gave a five-minute presentation on the paper at this year’s virtually held PKC.

“Now, I’m certain that simplicity is a big advantage as long as the result is useful or interesting.”

On June 24, ML@GT faculty members Srijan Kumar (School of Computational Science and Engineering,) Aditya Prakash (School of Computational Science and Engineering,) Munmun De Choudhury (School of Interactive Computing,) Nicoleta Serban (H. Milton Stewart School of Industrial and Systems Engineering,) and OMSCS student Kenneth Miller will participate in a virtual panel discussing their work. The panel will be moderated by ML@GT executive director Irfan Essa.

Panelists will give individual presentations before participating in a general question-and-answer segment with audience members.

• Kumar and De Choudhury will share details of their work regarding the psychological impact of Covid-19. Kumar will also discuss his work examining hate and counter-hate messages on Twitter against Asian Americans during the pandemic.
• Prakash is a member of the Center for Disease Control and Prevention’s (CDC) forecasting team, and will share their new data-driven approach to disease forecasting.
• Serban’s presentation will focus on her work creating an agent-based simulation forecasting model. This model captures the progression of the disease in an individual and in households, schools, communities, and workplaces.
• A lawyer by day and OMSCS student by night, Miller participated in a Kaggle challenge using natural language processing and machine learning to help doctors and scientists read the most important studies related to Covid-19.

The panel will take place virtually via a Bluejeans Event at 11 a.m. on June 24 and is open to the public. Registration is required.

The research makes program analysis simpler and more efficient by eliminating ineffective parts of the graph.

“InterDyck reachability is perhaps the most popular abstraction to perform interprocedural program analysis,” said Yuanbo Li, a Ph.D. student in SCS and co-author on the paper.

Many program analysis problems can be viewed as analyzing matching properties in graphs. InterDyck languages can help illuminate those properties. As a balanced-parentheses language, Dyck languages are often used to match one particular program property (such as function calls and returns), but the InterDyck language enables matching multiple balanced-parenthesis properties in program analysis simultaneously. Unfortunately, it's impossible to obtain the exact solution for InterDyck-reachability problems in general.

This research offers a framework to make the existing InterDyck-reachability algorithms more precise and scalable. If a graph edge isn’t contributing, it can be eliminated from the underlying input graph G.

“An edge is removable if and only if it won't affect the precise all-pairs reachability information, but computing the precise solution is undecidable,” Li said. “We are computing an over-approximation of InterDyck reachability. Therefore, we could safely identify a subset of removable edges.”

The researchers’ algorithm simplifies the input graphs and improves exisiting InterDyck reachability algorithms. After applying the simplification algorithm to a taint analysis for Android, all existing InterDyck-reachability algorithms ran much faster and became more precise. In the future, the researchers hope to apply the concept to even more general static analyses.

Li wrote the paper, Fast Graph Simplification for Interleaved Dyck-Reachability, with SCS Assistant Professor Qirun Zhang  and University of Wisconsin, Madison Professor Thomas Reps. They presented it at the 41st PLDI, an Association for Computing Machinery Special Interest Group on Programming Languages (ACM SIGPLAN) conference, held virtually from June 15 to 20.

This was one of three SCS papers at the conference. Zhang and Li presented another paper Debug Information Validation for Optimized Code, co-authored with SCS Ph.D. student Shuo Ding and Apple’s Davide Italiano. Professor Santosh Pande also presented BlankIt Library Debloating: Getting What You Want Instead of Cutting What You Don’t with SCS Ph.D. student co-authors Prithayan Barua, Girish Mururu, and Chris Porter.

To ensure Americans can find the most accurate information, College of Computing researchers are creating machine-learning (ML) and data science tools to help fact-checkers be more efficient.

The Disinformation Dilemma

Although having high-quality news is important any time, the ever-changing nature of COVID-19 makes it even more vital that users have access to vetted information. Many Americans receive their news from social media, where rumors can be shared as much as memes.

“Rumors, hoaxes, fake cures, bioweapon claims, and disinformation campaigns about COVID-19 are prevalent on social media,” said School of Computational Science and Engineering Assistant Professor Srijan Kumar. “These induce anger, anxiety, and stress in readers, and in many cases, have even led to fatalities, such as hydroxychloroquine overdose.”

Newsroom fact-checkers are at forefront of fighting against false information, but manually verifying every fact is time-consuming at best and nearly impossible in the age of COVID-19. So Kumar and School of Computer Science Professor Mustque Ahamad are building data-driven, secure solutions for fact checking.

A Next-Generation Solution

Kumar and Ahamad are a well-matched team. In his past cybersecurity research, Ahamad has worked with professional fact-checkers to determine what they need to complete their work at news organizations. Kumar, for his part, has been building ML and data-driven tools to detect disinformation. COVID-19 seemed like a natural pairing for the two.

“Together, we started collaborating to build the next generation of data-driven and security-minded solutions for effective fact checking,” Kumar said.

Their solution is to do early detection of disinformation before it even gets to the fact-checkers. With this in mind, they plan to develop ML techniques to remove deliberately misleading information from the news.

Their ML models will be able to learn the difference between true versus false information with only a few training data points.

“Our models will triage the cases that are most likely to be false in order of their impact on the readers,” Kumar said.

The models will also be customizable to the individual fact-checker’s topical, geographical, and language preferences. As the project develops, Kumar and Ahamad will collaborate with professional fact-checkers to ensure the models are effective throughout the research.

“Our framework will bring together a one-stop-shop for group of fact checkers to collaboratively identify false information,” Kumar said. “This information can then be shipped to appropriate stakeholders, so that the readers can be appropriately alerted when they view it and the hoaxes can be removed from social media circulation.”

For more coverage of Georgia Tech’s response to the coronavirus pandemic, please visit our Responding to COVID-19 page.

The program will fund at least 49 postdoctoral fellows in computing for a two-year career development experience. The National Science Foundation is funding the initial $14 million initiative.

Reviving the program

School of Computer Science Professor Ellen Zegura is the principal investigator (PI) on CIF 2020. As the chair of the CRA Board of Directors, Zegura said they anticipated that the COVID-19 pandemic and its economic fallout would impact graduating Ph.D. students, and have downstream effects on the U.S.’s computing research ecosystem.

“In late March and early April, CRA gathered data on how many academic departments were curtailing hiring to understand the scope of the problem,” she said. “Even at that time, this this led us to understand that many faculty positions were being pulled. We were wondering whether there was something CRA could be doing to be responsive to COVID-19.”

In response to the 2008 recession, CRA did something similar, funding more than 100 postdoctoral fellows and helping to popularize the fellowship model.

“This postdoctoral fellow idea bubbled up again because the circumstances were quite similar to 2008, but more drastic and quite sudden,” Zegura said.

This iteration of the program is also more flexible. Students can stay at their current institution but partner with a different advisor or work virtually with a mentor at a different university.

Matching researchers

Each candidate will identify a potential faculty mentor and can submit for the fellowship with up to two. There is a limit of two fellows and two mentors per institution to ensure this opportunity is spread across computing departments.

Although the NSF award has five PIs and a steering committee, there will also be a selection committee. These 8-10 faculty members broadly covering core computing research areas and will determine the matches.

Yet, according to Zegura, the goal is more than just bridging the job market gap. Each mentor and student pairing will plan a research topic and other ways the candidate can grow professionally, such as grant writing, teaching, or supervising junior students.

“This is supposed to be a time of growth,” Zegura said. “The postdoc experience really helps the person further the career.”

Application process

The application process is similar to faculty hiring processes to make it easy on applicants. Applicants will be required to submit: a research statement and plan, a mentoring plan written collaboratively with their proposed mentor, a curriculum vitae, and a few letters of recommendation. Students must be a U.S. citizen or permanent resident or have received a Ph.D. from a U.S. institution.

As with all CRA endeavors, increasing diversity in computing will be important in the selection process, Zegura said. The selection committee will consider under-represented minorities in computing —including women and racial and ethnical minorities — and geography, such as rural versus urban areas.

“This is a great example of what CRA can do well,” Zegura said. “We feel so good about being able to do something for the community and do something that’s responsive to COVID-19.”

The University System of Georgia and Georgia Tech recognizes tenured professors with this honor for their research excellence and contributions to the Institute. Zegura has made a career of networking and computing for social good research and teaching.

“Over my 27 years on the faculty at Georgia Tech, I have frequently been asked if I considered going elsewhere,” said Zegura, who is a Stephen P. Fleming Chair in Telecommunications. “The answer is that I have been able to do so many interesting and different things all while remaining right here. I am proud to have witnessed and contributed to the incredible growth of the university.”

Zegura is a leader across computing. She was named head of the Computing Research Association’s Board of Directors in 2018. She also leads the Center for Computing and Societies and its partnership with Georgia State around public interest technologies, a program that partners Georgia Tech computer scientists with Georgia State University social scientists to address social inequities.

Her research also connects networking and public interest. Last year, Zegura received a $150,000 Mozilla grant to incorporate ethics to the undergraduate computing curriculum. She has also researched how to bring better networks to Native American tribal lands.

Prasad Tetali, a joint professor in the Schools of Computer Science and Math, also had his Regents’ professorship renewed this year. Interim Chair Mostafa Ammar is also a Regents’ Professor, and Professor Seymour Goodman holds one joint with the Sam Nunn School of International Affairs.

"Being named Regents' Professor at Georgia Tech is a recognition of a faculty member's long-term distinguished scholarly and service contributions to their intellectual discipline, the academic profession and the Institute,” said Ammar. “I was very pleased to see Ellen Zegura receive such a well-deserved recognition. Prasad Tetali's renewal of his 2017 Regents' Professor appointment is further recognition of his continued impact and contributions. "

For the past 14 years, Google has recognized more than 100 faculty members across the U.S. for cutting-edge research in top computing areas from machine learning to quantum computing. The program funds one student worker and enables Google employees to collaborate with top academic institutions.

“I am particularly excited to receive this award,” Daglis said, “not only because it is the first form of direct support I’ve received from the industry, but also because I see it as a practical recognition of this research direction beyond academia, indicating a potential practical impact on datacenter systems in the near future.”

Daglis’s work focuses on using hardware specialization to make datacenters run more efficiently. By upgrading a datacenter network from a passive to an active system component, Daglis can leverage new relationships between incoming network traffic and CPU processing to improve processing efficiency, response latency, and ultimately create better online services.

Datacenters are fundamental for internet access because they allow millions of users access to online services at scale. Traditionally, thousands of central processing units (CPUs) communicated across a network to create a datacenter. As demand increased, datacenters  kept up by growing in size and continuously upgrading their CPUs, but this approach is now hitting performance scaling limitations because of the end of silicon scaling, such as Dennard scaling and slowdown of Moore’s law.

“CPU performance improvements alone are not sufficient anymore to keep up with booming user bases, exploding datasets, and new demanding services,” Daglis said. “This realization has led to the ongoing phenomenon of hardware specialization.”

With this in mind, Daglis’s research is to upgrade the datacenter network from a traditional passive approach that delivers data between two endpoints to an active component that performs smarter data manipulation. His emphasis is on the network’s endpoints – particularly the use of modern programmable network interfaces that can be leveraged to accelerate common inter-CPU communication patterns within the datacenter.

“The main underlying idea is that by judiciously exposing some application-level semantics to the network interface,” Daglis said, “new synergies between the incoming network traffic and the processing performed on the CPUs can emerge.”

Jake Abernethy has been promoted to associate professor in the School of Computer Science and granted tenure. Abernethy’s research focus is machine learning, where he enjoys discovering connections between optimization, statistics, and economics.

In 2011, he completed his Ph.D. at the University of California, Berkeley before becoming a Simons postdoctoral fellow for the following two years. After the water crisis in Flint, Mich., Abernethy worked on detecting lead contamination and infrastructure remediation. Prior to studying and teaching machine learning, Abernethy performed comedy and juggling shows, opening for Sinbad and Dave Chappelle.

Munmun De Choudhury has been promoted to associate professor in the School of Interactive Computing and granted tenure. De Choudhury is also affiliated with the GVU Center and Institute for People and Technology (IPaT) and leads the Social Dynamics and Wellbeing Lab (SocWeb Lab.) De Choudhury studies problems at the intersection of computer science and social media, building computational methods and artefacts to help understand human behaviors and psychological states and how they manifest online.

Prior to joining Georgia Tech in 2014, De Choudhury was a postdoctoral researcher in the nexus group at Microsoft Research, Redmond. In 2011, she received her Ph.D. from Arizona State University, Tempe. After graduate school, De Choudhury spent time at Rutgers University and was a faculty associate with the Berkman Center for Internet and Society at Harvard University.

Yajun Mei has been promoted to professor in the H. Milton Stewart School of Industrial and Systems Engineering. Mei's research interests include change-point problems and sequential analysis in mathematical statistics and sensor networks and information theory in engineering. Mei also examines longitudinal data analysis, random effects models, and clinical trials in biostatistics.

Mei received his Ph.D. in mathematics from the California Institute of Technology in 2003. He has also worked as a postdoc in biostatistics at the Fred Hutchinson Cancer Research Center. In 2010, Mei was awarded the National Science Foundation (NSF) CAREER Award and in 2008 was awarded Best Paper at FUSION. Mei was awarded the prestigious Abraham Wald Prize in Sequential Analysis in 2009.

Alex Endert has been promoted to associate professor and granted tenure in the School of Interactive Computing. Endert directs the Visual Analytics Lab where he and his students apply fundamental research to domains including text analysis, intelligence analysis, cybersecurity, and decision-making, and explore novel user interaction techniques for visual analytics.

Endert earned his Ph.D. from Virginia Tech in 2012, and in 2013 his work on Semantic Interaction was awarded the IEEE VGTC VPG Pioneers Group Doctoral Dissertation Award, and the Virginia Tech Computer Science Best Dissertation Award. In 2018, Endert received the NSF CAREER Award.

Editors Note: Molei Tao has been promoted to associate professor with tenure in the School of Math. Tao is an applied and computational mathematician, designing algorithms for faster and more accurate computations and developing mathematical tools to analyze and design engineering systems or answer scientific questions.

He earned his Ph.D. in control and dynamical systems with a minor in physics from the California Institute of Technolgy where he also worked as a postdoctoral researcher. He is the 2011 recipient of the W.P. Carey Ph.D. Prize in Applied Mathematics and a 2019 NSF CAREER Award.

The School of Computer Science (SCS) has promoted three faculty members. Hyesoon Kim and Santosh Pande became full professors, and Jacob Abernethy was granted tenure as an associate professor.

“Promotion and the award of tenure are some of the ways we as an academic community recognize the accomplishments of our people,” said Interim Chair Mostafa Ammar. “I am extremely pleased that we are able to recognize Jake, Hyesoon, and Santosh. They have always been valuable members of the SCS family, and I look forward to their continued success.”

Hyesoon Kim

Kim joined SCS in 2007 as a computer architect. She focuses on developing tools to improve performance and efficiency on heterogeneous architectures such as central processing units, graphics processing units, and embedded systems.

“I am very excited about the promotion; it means all lot to me,” Kim said. “When I was promoted to associate, I felt that I should stop running and look around because where else would I go. Now I feel that I would learn how to fly. “

Kim runs the High Performance Architecture Lab at Georgia Tech, where students and faculty develop high-performance and energy-efficient computing from microarchitectures to compilers.

“Thank you to all the people who helped me to build my career at Georgia Tech, including my graduate and undergraduate students, mentors, colleagues,” she said.

Santosh Pande

Pande has been with SCS since 2000. His work in compiler optimization is prolific and has considerable influence. Pande has published more than 100 papers, worked with top industry partners, and released several software systems

"This promotion means a lot to me looking back as well as peeking into the future,” Pande said. “I feel gratified for the recognition of scholarship in terms of research and education and its impact on the body of students I worked with at all levels.”

Pande focuses on practical problems he can provide foundational solutions for, such as evolving architectures for systems and security problems. With this in mind, Pande is currently working on solving software debloating problems and model-based program analysis and optimizations.

“I fell in love with compilers, which offered a unique blend of theory and practice and continue to be fascinated by its elegance (an art form), its deep theoretical moorings (science), and its robust engineering practices,” he said. “As I assume my new rank, I feel even more empowered and excited about making deep contributions to the advancement of my area and towards shaping up ambitious careers of my students at all levels."

Jacob Abernethy

Abernethy started at SCS in 2017 but has quickly made a mark with machine learning solutions to real-world problems such as the Flint Water Crisis. His work has received national news attention and National Science Foundation funding. He is also recognized as the leader in his field, working at Google Research in the Summer of 2019 and serving as program co-chair for the 2020 Conference on Learning Theory (COLT).

Abernethy has long felt part of the College of Computing and honored to receive this acknowledgment.

“In my first year in grad school, I watched [SCS Professor] Santosh Vempala give a phenomenal keynote talk at a conference I was attending,” he said. “I recall thinking at the time that becoming a tenured professor at Georgia Tech would be just an unimaginable accomplishment. It is amazing to think that 13 years later I would manage to get there myself!”

The rise of big data created research opportunities for new memory hardware that would offer higher capacity at a lower cost compared to traditional technologies. At the same time, the field of machine learning (ML) was booming.

Doudali saw a chance to have an influence on a new paradigm in computing and has built a research career on using ML-based methods to manage emerging memory systems.

All in the Family

Doudali has always taken advantage of the opportunities around her. Computing runs in the family. Her father is an electrical engineer, and her older brothers took an interest in computing as kids. Their passion meant Doudali got to play around on her first computer at 12 to explore YouTube and other media.

Despite her early exposure to computers, Doudali didn’t think about computer science as a field of study until her final two years of high school. With a plan to be a doctor, she was taking a lot of biology classes but noticed how much her brothers enjoyed their computer science studies at the National Technical University of Athens (NTUA) in Athens, Greece.

When Doudali started NTUA, she joined undergraduate program of electrical and computer engineering. Doudali completed a thesis on databases, creating a dataset generator for social networks enriched with geographic and textual information. This research experience at the Computing Systems Laboratory (CSLab) at NTUA showed Doudali where she wanted to take her career.

“I knew I wanted to stay in a university and wasn’t ready to go into industry yet,” she said. “I really liked the research process and collaboration.”

Research Agenda

A Ph.D. was the next step, and Doudali already knew about Georgia Tech thanks to her older brother, who completed a Ph.D. in computer architecture. She was instantly taken with the warm and welcoming environment of the school on her visit day.

The school embraced her, too. Doudali immediately started working with Associate Professor Ada Gavrilovska, who introduced Doudali to her future research field — heterogeneous memory systems.

Heterogeneous memory systems use new hardware technologies, such as persistent memory, or resource disaggregation techniques to meet the high capacity needs of big data applications at a reasonable cost compared to traditional systems. The varying access speeds of the available memory components make it necessary for intelligent data management methods to boost application performance.

“It was an interesting problem because heterogeneity now exists across the system’s hardware stack, like compute, memory, and storage units,” Doudali said.

A New Opportunity

As Intel’s Optane DC persistent memory rolled out, Doudali knew this was her window to do research that could really impact how a new product is used in real-time. Her research group harnessed ML to improve upon the challenging process of deciding where to place data in heterogeneous memory systems.

“We were interested in how ML could shape better systems,” Doudali said.

The result was Kleio, an ML-based data management methodology for heterogeneous memory systems. Kleio uses ML to predict data access pattern behavior and decide which data to move between the available memory units to improve application performance. The project has only continued to grow as more research questions open, such as how much data to move and at what times given network conditions. 

[Machine Learning Makes Memory Management More Efficient]

Doudali’s work has been recognized across the field. Kleio was a best paper award finalist at the   28th International Symposium on High-Performance Parallel and Distributed Computing (HPDC 2019). Recently, her alma mater invited her to give a talk on Kleio. She also served on a committee for the top systems conference EuroSys, which allows researchers to become familiar with the selection process of the technical program in academic conferences.

To boost her academic career, Doudali has acquired industry experience while interning at AMD, VMWare, and DellEMC. Ultimately, she plans to continue her research career after graduation.

Community Involvement

Doudali also gives back to the community she found a home in. She helps organize SCS visit days for prospective Ph.D. students and student meetings with faculty candidates. She is also an executive board member of the Hellenic Society, a cultural organization that connects students of Greek heritage and shares Greek traditions with the campus community.

 Doudali’s impact on the community has been appreciated. In April, she was invited to share four life quotes that shaped her Ph.D. journey as part of GT’s Illuminate Tech speaker series.

“Just let go of the things you cannot control right now,” she said during the event, “keep working hard, and trust that whatever happens is for the best.”

Associate Professor Taesoo Kim’s Systems Software & Security Lab (SSlab) Ph.D. students Insu Yun, Yong Hwi Jin, and Jung Won Lim competed in the annual event. Their objective was to exploit popular software with unknown vulnerabilities.

“As a hacker studying offensive security, we always dreamed of participating in Pwn2Own,” Yun said.

The SCS team exploited Apple’s Safari internet browser. Although the browser category is known to be more difficult, the team was attracted to how technically interesting the challenge was, according to Yun.

The team was also set up for success because the underlying Safari operating system is *nix, a category the team members were familiar with. Adding to their advantage, they also had written an exploit for Safari just last July so they were able to apply that knowledge here as well.

Multiple vulnerabilities are required to attack a modern browser. The team found new six vulnerabilities to compromise Safari, all of which were later confirmed by Apple. To exploit as many vulnerabilities as possible, they used several approaches, including fuzzing, source code review, and reverse engineering.

Ultimately, the competition allowed the team to sharpen skills they can bring back to their SCS research, such as designing automatic tools to find bugs, and identifying vulnerabilities on complex, real-world programs.

“SSlab gave us the opportunity to make a such wonderful team because our
lab is one of the best information security labs in the world with many talented students.”

This is why Georgia Tech researchers created a new tool called Desensitization that generates crash reports that preserve the original error — whether a bug or attack — without exposing privacy. The method is also smaller than half of all crash reports, taking less than 15 seconds to process and significantly saving resources for both users and developers.

“Crash dumps can have a severe outcome, including sensitive information leaking from end users, bad publicity, and financial liability for developers if a data breach happens,” said School of Computer Science Ph.D. student Ren Ding. “We came up with the idea to desensitize crashes while keeping necessary attack payloads before sharing it to the developers.”

Desensitization is effective at removing more than 80 percent of potential sensitive data from Linux crash reports and nearly 50 percent of Windows crash reports.

The problem with crash reports

Crash reports usually include a coredump file that contains the central processing unit context and memory of the program, or the program inputs that made it crash. Both can include sensitive data, from session tokens to personal contact information.

Previous techniques to remove this type of data weren’t effective. Relying on developers to hand-annotate the data is time consuming and error prone. Another method that uses a pattern-based search to identify private data, such as email addresses, doesn’t work on program-specific data. The biggest issue with both methods is how much computation they require.

How Desensitization works

In contrast, Desensitization first runs on the user side to decouple general information and crash evidence from personal data. The order is strategic, according to Ding. End users may not have as much computation power as developers, so debugging and analysis isn’t efficient at the source.

It’s imperative to minimize necessary resources such as time and memory, making the program “lightweight.” The goal, said Ding, is to extract enough general information so developers can conduct the crash diagnostics on their end. Once the tool extracts the attack information from the user, it’s sent to the server for more thorough analysis.

The researchers’ method is bug and attack oriented to ensure the tool only focuses on relevant data. With this in mind, they designed one general and four lightweight techniques.

The general technique scans the memory to identify all pointers, a programming language object that stores memory and where most attacks reside. The four other techniques either identify specific information or customize existing techniques in prominent bugs and attacks.

“Our security model relies on the security guarantee of previous detection methods,” Ding said. “That’s why we try to make the current framework customizable by adopting the module design so that it can support more in the future.”

The module design of Desensitization makes the tool easy to apply to future technique and works on existing crash report analysis formats.

The solution is available to developers via Github.

Ding will present the research at the Network and Distributed System Security Symposium (NDSS) in San Diego running from Feb. 23 through 26. He co-wrote the paper, Desensitization: Privacy-Aware and Attack-Preserving Crash Reports, with fellow SCS Ph.D. students Hong Hu and Wen Xu and Associate Professor Taesoo Kim.

As the cloud has enabled more complex applications everywhere, some of the most vital computing occurs at the edge of the network, from video processing to image recognition. Yet with these innovations come challenges, including latency, cost, and security. This is where SCS steps up.

Across the school, networking and systems researchers tackle how best to take advantage of the edge’s many benefits while ensuring it’s secure and practical. Together they are shaping the future of the field.

In this video, several SCS professors, research scientists, and students explain the power of the edge, their research, and what to expect next. Interim Chair Mostafa Ammar, Professor

Ellen Zegura, Associate Professor Ada Gavrilovska, Research Scientist II Ketan Bhardwaj, Professor Umakishore Ramachandran, and Ph.D. student Harshit Gupta show how we are dominating in this field.

“In the past five years, researchers are showing a lot of different vulnerabilities in hardware,” School of Computer Science Ph.D. student Nader Sehatbakhsh said. “Power management units are just one of many different hardware components that can leak secrets.”

The Power Problem

While computer power used to be a simple on-off switch, new PMUs operate in up to 10 different states to preserve energy. Researchers found that each state emits a different electromagnetic signal that creates spikes in the frequency-domain spectrum, and as the activity increases so does the signal and the spikes.

Typing characters creates periods of high-state activities in which the distance between state spikes can show the location of keys on a keyboard. With the right monitoring, an attacker can determine the number of characters or how many words are in a password, or even find the exact characters.

This type of monitoring is called a side-channel attack, in which an attacker can extract sensitive data from physical signals produced by electronic activity within the device. Sound, temperature, power, and electromagnetic waves are common signals.

[RELATED: Researchers Help Close Security Hole in Popular Encryption Software]

PMU side-channel attacks can also be exploited to create a fast and stealthy covert communication. For example, a rogue employee could extract data outside of a secured company’s computer, breaking existing security.

Stopping the Side-channel

Side-channel vulnerabilities are becoming more pervasive because security is not the first priority for hardware designers, according to Sehatbakhsh. Their primary goals are to create something fast and cost-effective, and security often slows down the process and becomes more expensive. Yet with more research like this, things could change, Sehatbakhsh said.

“Now people are demanding extra security, so hopefully ignoring these vulnerabilities won’t be an option in the future.”

At this point in their work, the researchers are mostly exposing the problem, but there are potential solutions, Sehatbakhsh suggested. Adding randomness to the PMU would unsync power states from the data they provide; however, this solution might consume too much power for developers to implement it.

Sehatbakhsh will present the research at the 26th International Symposium on High-Performance Computer Architecture in San Diego from Feb. 22 to 26. He co-wrote the paper with School of Electrical and Computer Engineering (ECE) Ph.D. student Baki Berkay Yilmaz, ECE Associate Professor Alenka Zajic, and SCS Professor Milos Prvulovic.

• MS18 Exploiting Task Parallelism in Exascale Computing Era
• MS65 High-Performance Tensor Computation and Applications - Part II of III
• MS51, MS62, MS72 Novel Computational Algorithms for Future Computing Platforms - Part III of III
• CP14 HPC for Data Science and Large Graphs 

Click the link to view Georgia Tech's complete participation at SIAM PP20.

Laddha is a second-year student in the Algorithms, Combinatorics and Optimization program, where she specializes in creating faster algorithms for high-dimensional sampling and convex optimization. This work is foundational in improving machine learning.

The fellowship funds three years of Laddha’s Ph.D. and provides the opportunity to visit Microsoft Research for a two-day summit.

“This fellowship means a lot to me,” Laddha said. “It’s a very good opportunity to see the kind of research carried out in the industry.”

The application process for the fellowship was competitive. Laddha was nominated by her advisor, Professor Santosh Vempala, and then did an in-person interview with Microsoft.

“Aditi Laddha is working on central problems in algorithms, specifically sampling and optimization in high dimension,” Vempala said.  “Her clear thinking, raw mathematical talent, and creativity make her a richly deserving recipient of the Ada Lovelace fellowship.”

"It is an honor to become an IEEE fellow, and always a great feeling to be recognized by one's colleagues," said Sarkar, who is the Stephen Fleming Chair for Telecommunication in the College of Computing and co-director of the Center for Research into Novel Computing Hierarchies (CRNCH).

As the world’s leading professional association for advancing technology for humanity, this IEEE distinction is considered to be a prestigious honor and an important career achievement in the technical community. This year the IEEE Board of Directors honored three GT faculty members: Sarkar, School of Computational Science and Engineering Professor Richard Fujimoto, and School of Electrical and Computer Engineering (ECE) Professor Stanislav Emelianov.

[RELATED NEWS: Fujimoto and Sarkar Named IEEE Fellows]

Sarkar was recognized for contributions to high-performance computing compiler technologies. As a parallel computing expert, Sarkar’s work includes programming languages, compilers, runtime systems, and debugging and verification systems for high performance computers.

This is not Sarkar’s first honor in his field. He became a member of the IBM Academy of Technology in 1995 and an Association of Computing Machinery (ACM) Fellow in 2008. He has been serving as a member of the U.S. Department of Energy’s Advanced Scientific Computing Advisory Committee (ASCAC) since 2009 and has served on Computing Research Association’s Board of Directors since 2015. 

Sarkar is the sixth IEEE fellow in SCS. He joins Interim Chair Mostafa Ammar, CRNCH Co-Director Tom Conte, Professors Ling Liu, Umakishore Ramachandran, and Ellen Zegura.

"In many ways, the synergies between ACM and IEEE is reflected in the synergies between the College of Computing and the School of Electrical and Computer Engineering at Georgia Tech, and can be seen in multiple close research collaborations on campus including in our CRNCH center,” said Sarkar.

This annual gathering has become one of the top forums to discuss the future of computing after Moore’s law, the past industry-wide trend of doubling transistors in a microchip nearly every two years that exponentially fueled computing innovation. Now dozens of experts will meet to discuss computing’s new frontier from the perspective of diverse areas including devices, edge computing, computer architecture, systems software, machine learning, quantum computing, and theory.

“The CRNCH Summit is an exciting opportunity for Georgia Tech researchers and our visitors to share their latest breakthrough ideas for post-Moore computing,” said CRNCH Co-Director Vivek Sarkar.  “We look forward to a productive meeting, and all the great research that will follow.”

This year’s keynote is on computer architecture by David Mountain, the senior technical director at Advanced Computing Systems Research Program. Other leaders in their field follow, including Oak Ridge National Laboratory AI Institute Director David Womble, Notre Dame Professor Peter Kogge, National Instruments Academic Business Development Manager Igor Alvarado, Northrop Grumman System Architect Brian Konigsburg, and GTRI Quantum Systems Division Senior Research Scientist Craig Clark.

The event is also a chance for Georgia Tech faculty to showcase their research. School of Computer Science (SCS) Associate Professor Ada Gavrilovska discusses edge computing possibilities. SCS Professor Dana Randall offers emergent computation as a possibility. SCS Associate Professor Hyesoon Kim presents heterogeneous computing systems. School of Electrical and Computer Engineering Professors Arijit Raychowdhury and Moin Qureshi bring their perspectives on future devices and quantum computers.

The event highlights how CRNCH’s influence spans the entire institute with more than 30 dedicated faculty members from the College of Computing, the College of Engineering, the College of Sciences, and the Georgia Tech Research Institute. With their expertise in quantum computing, neuromorphic computing, design science, approximate computing, and more, CRNCH’s team is uniquely qualified to tackle the challenges of computing’s future. Since it was founded in 2017 by Co-director Tom Conte, the center has paired researchers with funding, students with internships, companies with research labs on campus to test their leading-edge products, and even started a collection of specialized hardware called the Rogues Gallery (RG).

The RG has supported close to 75 users with a quarter coming from external labs and institutions. In 2019, co-directors Jason Riedy and Jeff Young also ran external tutorials at the ASPLOS and PEARC conferences and presented RG-related work at multiple venues including SIAM CSE, PEARC, and ICRC. Looking forward, Young expects more challenging projects.

 “In 2020, the Rogues Gallery will see the deployment of additional hardware focused on Arm high-performance computing as well as new neuromorphic prototypes and support for quantum programming” he said. “We’re looking to support a more diverse set of research hardware as well as a more inclusive userbase to tackle the toughest challenges in post-Moore computing”.

Emelianov was recognized for his contributions to ultrasound elasticity and photoacoustic imaging. He is the Joseph M. Pettit Chair Professor in the School of Electrical and Computer Engineering and a Georgia Research Alliance Eminent Scholar. An expert in biomedical imaging instrumentation and nanoagents for imaging and therapy, Emelianov has joint appointments with the Wallace H. Coulter Department of Biomedical Engineering at Georgia Tech and Emory University. He is also a professor of Radiology at the Emory University School of Medicine and is affiliated with Winship Cancer Institute and other clinical units. 

Emelianov is the director of the Ultrasound Imaging and Therapeutics Research Laboratory, where his group works on the discovery, development, and clinical translation of diagnostic imaging and therapeutic instrumentation, augmented with theranostic nanoagents–small particles that can diagnose and then treat a specific disease. He is a Fellow of the American Institute for Medical and Biological Engineering, and he has served as vice president for Ultrasonics of the IEEE Ultrasonics, Ferroelectrics, and Frequency Control Society.

Fujimoto, a Regents’ Professor in the School of Computational Science and Engineering, was honored for his work in the field of parallel and distributed discrete event simulation. Discrete event simulations model operations within a system and have uses in a wide variety of applications. Fujimoto has authored and co-authored hundreds of technical papers on the subject as well as several books, which span application areas including transportation systems, telecommunication networks, and multiprocessor and defense systems.

He was also named a 2019 Interservice/Industry Training, Simulation and Education Conference (I/ITSEC) Fellow. The announcement for both of these recognitions came only two years after he was named an Association for Computing Machinery Fellow in 2017.

Sarkar, the Stephen P. Fleming Chair of Telecommunications in the School of Computer Science and co-director of the Center for Research into Novel Computing Hierarchies, received his distinction for contributions to compiler technologies for high-performance computing. His work in this area spans multiple aspects of parallel computing software including programming languages, compilers, runtime systems, and debugging and verification systems for high performance computers.

Sarkar has numerous recognitions in the field. He became a member of the IBM Academy of Technology in 1995 and an ACM Fellow in 2008. He has been serving as a member of the U.S. Department of Energy’s Advanced Scientific Computing Advisory Committee (ASCAC) since 2009 and has served on CRA’s Board of Directors since 2015. 

The IEEE – short for the Institute of Electrical and Electronics Engineers – is the world’s leading professional association for advancing technology for humanity. Through its 420,000-plus members in more than 160 countries, the association is a leading authority on a wide variety of areas ranging from aerospace systems, computers and telecommunications, biomedical engineering, electric power, and consumer electronics.

Dedicated to the advancement of technology, the IEEE publishes nearly one-third of the world’s literature in the electrical and electronics engineering and computer science fields, and has developed nearly 1,300 active industry standards.  The association also sponsors or co-sponsors more than 1,900 international technical conferences and events each year.