From Robustness to Privacy and Back
summary
Gives a black-box transformation that turns any robust estimator into an ε‑differentially private estimator, with the DP error matching the robust error for a specific corruption level determined by parameter‑space size, sample size and ε. Shows the construction is tight: DP procedures produced are optimal in low dimensions and optimal for several high‑dimensional tasks if the robust estimator is optimal. Establishes an equivalence between ε‑privacy and robustness at a corruption scale about log n/ε.
AI-written summary · about
related talks
17mData privacy and reliable learning
28mPrivacy versus Robustness in Federated Learning: Limits and Algorithms
43mDifferential Privacy Meets Robust Statistics
21mDimension-free Private Mean Estimation for Anisotropic Distributions
01h 04mAlgorithmic High Dimensional Robust Statistics II
40mDifferential Privacy with Elliptical Stochastic Processes
26mAaron Roth, Associate Professor, University of Pennsylvania, NYC 2017
19mPractical Considerations for Differential Privacy