Accuracy First: Selecting a Differential Privacy Level for Accuracy-Constrained...
summary
Presents an accuracy-first procedure for differentially private ERM that searches epsilons by starting very private and doubling privacy while using correlated Laplace noise (continuous random walk) so earlier trials cost nothing. Adds a modified above-threshold check for adaptive, data-dependent queries and frames guarantees as ex post differential privacy. Empirically it beats flipping worst-case utility theorems and naive doubling: on ridge and logistic tasks (Twitter popularity, KDD99) observed privacy loss drops from ~500 to ~10 at alpha=0.05. Surprising find: checking accuracy dominates privacy cost; method fails to cleanly extend to (epsilon,delta).
AI-written summary · about
chapters (7)
- 0:00Introduction to DP
- 3:52Stability and mechanisms
- 10:16Accuracy-first objectives
- 16:09Principled epsilon search
- 20:24Correlated noise techniques
- 23:25Empirical performance results
- 26:53Challenges and conclusions
related talks
01h 00mDifferential Privacy: Fundamentals to Forefront I
35mDifferential Privacy Made Practical
42mDifferential Privacy
19mPractical Considerations for Differential Privacy
43mIs Interaction Necessary for Distributed Private Learning?
01h 05mMarkets and Incentives for Transacting on (Personal) Information, Part III
17mData privacy and reliable learning
27mPrivacy-preserving statistical analyses