Title: Systems Ascribing Privacy to Optimization Methods

Date: Wednesday, April 16th, 2025

Time: 12:30 PM

Location: Klaus Room 3402

James Choncholas

Computer Science PhD Student

School of Computer Science

Georgia Institute of Technology

Committee:

Dr. Ada Gavrilovska (Advisor) - School of Computer Science, Georgia Institute of Technology

Dr. Ketan Bhardwaj - School of Computer Science, Georgia Institute of Technology

Dr. Vladimir Kolesnikov - School of Cybersecurity and Privacy, Georgia Institute of Technology

Dr. Frank Li - School of Electrical and Computer Engineering, Georgia Institute of Technology

Dr. Ling Liu - School of Computer Science, Georgia Institute of Technology

Abstract:

Contemporary data analysis tasks, frequently formulated as optimization problems, face a critical efficiency challenge when privacy requirements prevent sharing data. Secure computation techniques like secure multiparty computation and homomorphic encryption are common candidates to introduce privacy, but their substantial performance overheads often make direct application impractical to optimization. This thesis explores how bespoke co-design of system architecture and cryptographic protocols achieves both practical performance and meaningful privacy guarantees in optimization tasks by exposing trade-offs between solution optimality and efficiency.

This work presents three novel systems: Angler for decentralized, dark pool resource allocation, posing only 2x overhead compared to non-private solutions; Snail for multi-party visual localization, reducing computational overhead by two orders of magnitude compared to a straightforward private baseline; and Hadal, an extensible data-flow framework for privacy-preserving machine learning introducing homomorphic encryption and cryptography-aware optimizations to TensorFlow. Together, these contributions navigate new efficiency-accuracy trade-off frontiers while maintaining privacy guarantees, making secure optimization practical for real-world applications.