Towards Understanding Privacy Leakage in Decentralized and Collaborative Learning

Abstract

The emergence of large-scale machine learning (ML) models has highlighted a fundamental conflict: While computational demands push for the consolidation of data and models in vast, centralized data centers, real-world data continues to be distributed and fragmented across personal devices and private databases. How can we reconcile this contradiction without further monopolizing the ML ecosystem? What unique privacy and security risks arise from alternative ML orchestration system designs? Furthermore, how do these vulnerabilities and system failures inform our understanding of both how and what machines learn? This thesis attempts to explore these questions. It first examines key types of privacy leakages, evaluating their impact under realistic, cross-distribution settings. It then introduces a benchmarking analysis platform, SONAR, to investigate the relationship between privacy leakage (measured by attack performance), network topology, and data distribution. Finally, it presents Co-Dream, a novel algorithm for collaborative learning that offers improved privacy characteristics.