Differential Privacy Meets Robust Statistics
Talkat Algorithmic Advances for Statistical Inference with Combinatorial Structure · 2021-10-11..10-15·
summary
An efficient algorithm that provably combines robustness to an α fraction of poisoned samples and (ε,δ)-differential privacy is presented, cutting iterative filtering from O(d) to O(log^2 d) by using matrix multiplicative weights to produce matrix-exponential “quantum” scores and a single-threshold private filtering (sample a uniform Z and remove all scores > Z) that couples randomized removals for privacy. Error incurs an extra ≈ d^{1.5}/(ε n) term (open whether tight). An exponential-time high-dimensional propose-test-release procedure achieves optimal private, robust rates for mean, covariance, PCA and regression and empirically removes backdoor attacks on CIFAR-10.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:28Membership Inference Attack
- 2:47Trust Issues
- 3:03Backdoor
- 3:40Experiment
- 7:45Motivation
- 10:06Robust Mean Estimation
- 16:36Making it Private
- 17:33Quantum Robust Mean Estimation
- 18:55Differential Privacy
- 27:24HighDimensional Proposed Test Release
- 28:05General Framework
- 34:00Proposed Test Release
- 35:29Modular Approach
- 37:00Practical Settings
- 39:02Questions
related talks
30mAccuracy First: Selecting a Differential Privacy Level for Accuracy-Constrained...
01h 00mDifferential Privacy: Fundamentals to Forefront I
09mFrom Robustness to Privacy and Back
17mData privacy and reliable learning
28mPrivacy versus Robustness in Federated Learning: Limits and Algorithms
26mGeneralized Private Selection and Testing with High Confidence
01h 04mAlgorithmic High Dimensional Robust Statistics II
47mA Blackbox Reduction for Adaptive Adversaries using Differential Privacy