General-Purpose f-DP Estimation and Auditing in a Black-Box Setting

Learn about novel statistical methods for assessing f-Differential Privacy (f-DP) through this 19-minute conference presentation from USENIX Security '25. Explore how researchers from Ruhr-University Bochum, Aarhus University, University of Victoria, and Georgia Institute of Technology address the challenge of validating f-DP mechanisms in black-box settings without requiring prior knowledge of the investigated algorithm. Discover new black-box methods that provide complete estimates of f-DP trade-off curves with theoretical convergence guarantees, and examine an efficient auditing approach that empirically detects f-DP violations with statistical certainty by combining non-parametric estimation and optimal classification theory. Understand how f-DP serves as a refinement of standard differential privacy, addressing weaknesses including tightness under algorithmic composition, and see experimental demonstrations of the effectiveness of these estimation and auditing procedures across various DP mechanisms.