Smooth Sensitivity, Revisited
summary
A differentially private Gaussian mean estimator that matches the nonprivate variance up to a low‑order privacy cost, achieved by applying smooth sensitivity to the trimmed mean and adding new noise families (notably the Laplace log-normal, plus Student‑t and sinh‑normal variants). The paper gives a computable smooth‑sensitivity G, MSE = optimal MSE + privacy term that decays with n, and experiments (n≈200–1000) show the best noise yields excess variance ~10% over nonprivate estimation.
AI-written summary · about
chapters (19)
- 0:00Intro
- 0:13Talk Outline
- 1:14Gaussian Mean Estimation
- 2:20Preview: Algorithm for Gaussian Mean
- 3:46Non-Privately: Empirical Mean
- 4:49New Approach: Trimmed Mean
- 7:07Trimmed Mean for Non-Gaussians
- 9:16Sensitivity of Trimmed Mean?
- 10:23Idea: Smooth Sensitivity NRS07
- 11:30Smooth Sensitivity Nissim-Raskhodnikova-Smith07
- 13:00Additive and multiplicative distortions
- 13:19Smooth Sensitivity Distributions NRS07
- 15:21New Smooth Sensitivity Distributions
- 18:07Aside: Concentrated Differential Privacy
- 19:17Laplace-logNormal Privacy Analysis
- 29:21Smooth Sensitivity Algorithm
- 32:44Smooth Algorithm for Gaussian Mean
- 33:45Smooth Sensitivity of Trimmed Mean Smooth sensitivity
- 43:08Conclusion
related talks
01h 14mTutorial: Privacy, Part II
31mGaussian Differential Privacy (GDP)
01h 02mLarge-Scale Private Learning, Part I
01h 07mLarge-Scale Private Learning, Part II
30mAccuracy First: Selecting a Differential Privacy Level for Accuracy-Constrained...
36mPrivately Learning High-Dimensional Distributions
17mData privacy and reliable learning
21mDimension-free Private Mean Estimation for Anisotropic Distributions