General-Purpose f-DP Estimation and Auditing in a Black-Box Setting
summary
Black-box procedure estimates the f-DP trade-off curve by sampling outputs from neighboring databases, fitting densities with a kernel density estimator, and emulating the likelihood-ratio test via a perturbed likelihood-ratio test (adding a small uniform perturbation h·u). Empirical results show accurate recovery with moderate samples (examples at n=200 and n=5,000) and uniform convergence guarantees. An confidence region built from optimal-classification error estimates flags provable f-DP violations on tested mechanisms.
AI-written summary · about
chapters (6)
- 0:00f-DP estimation and auditing intro
- 0:59f-differential privacy basics
- 7:21Why f-differential privacy is important
- 8:05Estimating f-DP in black-box
- 13:26Auditing f-DP claims
- 18:23Summary
related talks
26mGeneralized Private Selection and Testing with High Confidence
37mUser-level and federated local differential privacy
19mPractical Considerations for Differential Privacy
01h 14mStatistical Inference and Privacy, Part III
09mFrom Robustness to Privacy and Back
01h 03mPrivately Evaluating Untrusted Black-Box Functions
47mBlack-box Random Differential Privacy with Sensitivity Sampling
44mDifferentially Private Inference for Regression Coefficients