Private Meta-Selection
summary
Defines private meta‑selection: pick the best among K differentially private candidate mechanisms while avoiding an ε cost linear in K. Gives a randomized algorithm that samples the number of trials from a geometric distribution, outputs a model plus a noisy score, and provably competes with the best candidate median. Guarantees include 3×ε_base privacy, runtime ~K/β for target percentile P≈1/K, and success prob ≥1−β. Variants handle (ε,δ) bases and bounded runtime.
AI-written summary · about
chapters (8)
- 0:00Start
- 0:34Selection Problem
- 2:30Motivating the Problem
- 8:10The Meta Selection Problem
- 22:07Optimization
- 26:51History
- 30:51Private Reputation Theorem
- 41:32Open Questions
related talks
26mGeneralized Private Selection and Testing with High Confidence
34mOptimal Privacy-Constrained Mechanisms
30mAccuracy First: Selecting a Differential Privacy Level for Accuracy-Constrained...
41mpMSE Mechanism: Differentially Private Synthetic Data with Maximal Distributional Similarity
01h 02mLarge-Scale Private Learning, Part I
26mAlgorithms with More Granular Differential Privacy Guarantees
35mPrivacy amplification by random allocation
01h 03mStatistical Inference and Privacy, Part II