Tutorial: Privacy, Part II
summary
Defines differential privacy as an output-distribution bound parametrized by epsilon/delta, proves noise calibrated to global sensitivity (e.g., mean of bounded d‑dimensional points needs noise giving error ~sqrt(d)/n), and gives the practical recipe: clip per‑example gradients to C, add Gaussian noise to the averaged gradient, then use privacy amplification by subsampling so DP‑SGD is feasible. Covers composition/group‑privacy, moments accountant/Gaussian DP for tighter accounting, and practical caveats about clipping and neighbor definitions.
AI-written summary · about
chapters (9)
- 0:00Differential privacy intro
- 1:30Formal definition
- 5:58Noise addition for privacy
- 16:25Limits of differential privacy
- 20:18Composition and group privacy
- 37:53DP training overview
- 40:40DP-SGD algorithm
- 59:39Improving DP-SGD accounting
- 1:03:47DP optimization open questions
related talks
35mDifferential Privacy Made Practical
19mPractical Considerations for Differential Privacy
50mPrivacy Amplification by Sampling and Renyi Differential Privacy
31mGaussian Differential Privacy (GDP)
26mAaron Roth, Associate Professor, University of Pennsylvania, NYC 2017
31mWhen Existing Techniques in Linear Regression Preserve Differential Privacy
01h 00mPANEL: Statistical Theory, Privacy and Data Analysis
01h 07mStatistical Inference and Privacy, Part I