Black-box Random Differential Privacy with Sensitivity Sampling
summary
Paper proposes a sensitivity sampler that estimates the distribution of a function’s local sensitivities by drawing many i.i.d. samples, inverting the empirical CDF with the DKW inequality, and choosing a conservative sensitivity quantile to run standard mechanisms; guarantee is random differential privacy (confidence γ), not worst‑case ε‑DP, and achieving γ≈0.01 needs thousands–ten‑thousands of samples. Also implements a Bernstein mechanism for private function release (basis from Bernstein polynomials) and ships code as a CRAN package; demos work but require large m for strong confidence.
AI-written summary · about
chapters (10)
- 0:00Start
- 0:04Talk outline
- 4:30"The laws of mathematics are very commendable but the only law that applies in Australia is the law of Australia." -The PM
- 9:17Just bound sensitivity
- 14:21Idea 1: Sensitivity-induced privacy wlog --DP
- 36:17Example: Laplace for Linear SVM
- 37:05Bernstein Mechanism: Sketch
- 39:54Aside: Bernstein Function Approximation
- 40:28Bernstein Utility
- 42:42Example: Priestly-Chao kernel regression
related talks
01h 14mTutorial: Privacy, Part II
35mPrivacy amplification by random allocation
43mPrivacy Amplification by Sampling in Practice
19mPractical Considerations for Differential Privacy
47mDifferential Privacy from a Statistical Perspective–Obtaining Valid Inferences...
26mGeneralized Private Selection and Testing with High Confidence
01h 03mStatistical Inference and Privacy, Part II
01h 03mSimons Institute Live Stream