Privately Evaluating Untrusted Black-Box Functions
summary
Two settings: automated sensitivity detection (analyst supplies only black-box f and its finite range) and the provided sensitivity bound case. The work builds an epsilon‑differentially private “privacy wrapper” that queries f on x and its subsets, uses the shifted inverse mechanism (exponential mechanism / generalized interior point tools) plus a monotonization operator and Laplace steps, achieves near‑optimal accuracy/error ≈ c/ε and query complexity |x|^λ with λ matching proven lower bounds, and proves tight impossibility results.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 10:09Problem setting
- 20:58Differential privacy basics
- 31:30Quantitative results
- 40:58Privacy wrapper for sensitivity
- 56:04Monotonization and functions
- 1:01:26Summary
related talks
26mGeneralized Private Selection and Testing with High Confidence
30mAccuracy First: Selecting a Differential Privacy Level for Accuracy-Constrained...
47mBlack-box Random Differential Privacy with Sensitivity Sampling
19mGeneral-Purpose f-DP Estimation and Auditing in a Black-Box Setting
01h 00mDifferential Privacy: Fundamentals to Forefront I
45mMaking Privacy Technology Accessible: Benchmarks and Platforms
09mFrom Robustness to Privacy and Back
17mData privacy and reliable learning