Date: Monday, November 25, 2024

Time:  3:00pm-5:00pm (Eastern Time)

Location: Klaus, Room 3402

James Choncholas

School of Computer Science

College of Computing

Georgia Tech

Committee

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

Dr. Ling Liu - School of Computer Science, Georgia Tech

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

Abstract

Optimization tasks face a critical efficiency challenge when privacy requirements prevent sharing data held by multiple parties. Techniques such as secure multiparty computation and homomorphic encryption are candidates to introduce privacy, but their performance overhead can make direct application impractical. This thesis explores approaches to balance solution optimality in optimization with efficiency, coupling system design with cryptographic protocols to achieve sufficient utility, practical performance, and meaningful privacy.

This work presents three systems. Angler, performing decentralized, privacy-preserving resource allocation posing only 2x overhead compared to non-private solutions. Snail, which supports multiparty visual localization reducing computational overhead by two orders of magnitude compared to a straightforward private baseline. And Hadal, a dataflow-based framework for privacy-preserving machine learning introducing homomorphic encryption and cryptography-aware optimizations to TensorFlow. These systems collectively offer generalizable principles to ascribe privacy to optimization methods, making secure optimization feasible in real-world settings.