Gaussian Differential Privacy (GDP)
summary
Proposes Gaussian Differential Privacy as a replacement for single (ε,δ) pairs, defining privacy by a trade-off function between type I and II errors and then fixing that function to the Gaussian curve G_mu to get Gaussian Differential Privacy. Proves tight subsampling amplification, composes privacy by a tensor product that yields √k scaling for k i.i.d. μ-GDP mechanisms, and gives a central limit theorem showing small, properly scaled leaks converge to GDP, with improved bounds for moment-accountant-style analyses.
AI-written summary · about
chapters (12)
- 0:00Intro
- 0:35Controversial
- 1:52Basics
- 2:40Hypothesis Testing
- 3:28Smallest Type 2 Error
- 4:59Larry Weisman
- 8:30All Values of L
- 9:52GDP
- 12:11Postprocessing
- 16:06Bridging equality
- 21:21Theorem
- 24:41Central Limit Theorem
related talks
01h 14mTutorial: Privacy, Part II
26mAlgorithms with More Granular Differential Privacy Guarantees
35mDifferential Privacy Made Practical
50mPrivacy Amplification by Sampling and Renyi Differential Privacy
01h 03mStatistical Inference and Privacy, Part II
01h 02mLarge-Scale Private Learning, Part I
19mPractical Considerations for Differential Privacy
31mWhen Existing Techniques in Linear Regression Preserve Differential Privacy