Data privacy and reliable learning
summary
Shows that private learners can achieve dimension-free error for some mean estimation tasks by using instance-adaptive bounds and exploiting low effective rank, while proving any robust estimator converts to a differentially private one with an extra error term roughly D log R / (ε n); also proves differential privacy implies robustness to about a 1/(ε n) fraction of corruptions. Reduction relies on parameter-space coverings and robustness-to-privacy transformations and highlights when the covering blows up the dimension dependence.
AI-written summary · about
chapters (6)
- 0:00Introduction
- 1:03Differential privacy in learning
- 4:19Privacy and robustness links
- 7:13Making robust estimators private
- 9:23Instance optimality in learning
- 11:54Q&A
related talks
09mFrom Robustness to Privacy and Back
53mOptimality in Locally Private Estimation and Learning
43mIs Interaction Necessary for Distributed Private Learning?
01h 02mLarge-Scale Private Learning, Part I
28mPrivacy versus Robustness in Federated Learning: Limits and Algorithms
58mLarge-Scale Private Learning, Part III
30mAccuracy First: Selecting a Differential Privacy Level for Accuracy-Constrained...
43mDifferential Privacy Meets Robust Statistics