Randomized Approach for Tight Privacy Accounting in Differential Privacy

Explore a Google TechTalk presented by Jiachen T. Wang from Princeton University on a novel differential privacy paradigm called estimate-verify-release (EVR). Delve into the challenges of privacy accounting in differential privacy (DP) compositions and discover how EVR addresses these issues by converting privacy parameter estimates into formal guarantees. Learn about the core component of EVR, privacy verification, and the development of a randomized privacy verifier using Monte Carlo (MC) technique. Examine the proposed MC-based DP accountant and its advantages over existing DP accounting techniques in terms of accuracy and efficiency. Gain insights into how this new approach improves the utility-privacy tradeoff for privacy-preserving machine learning through empirical evaluation results.

Randomized Approach for Tight Privacy Accounting