Bridging Privacy and Robustness for Trustworthy Machine Learning

Abstract

The advent of machine learning has led to transformative changes across various domains, but the sensitive nature of data raises concerns about privacy and security. While Local Differential Privacy (LDP) has been a cornerstone in addressing these concerns, recent research has proposed privacy concepts aligned with the Bayesian inference perspective of an adversary, such as Average Bayesian Privacy (ABP) and Maximum Bayesian Privacy (MBP). This paper explores the intricate relationships between LDP, ABP, and MBP, and their implications for algorithmic robustness. We establish theoretical connections between these privacy notions, proving that LDP implies MBP and vice versa under certain conditions, and deriving bounds connecting MBP and ABP. We also investigate the relationship between PAC robust learning and privacy preservation, demonstrating how to derive PAC robustness from privacy-preserving algorithms and construct privacy-preserving algorithms from PAC robust ones. Our findings provide valuable insights for constructing privacy-preserving and robust machine learning algorithms.

Tasks

Reproductions