Differential Privacy from a Statistical Perspective–Obtaining Valid Inferences...
Talkat Data Privacy: From Foundations to Applications (Simons Institute workshop, 2019) · 2019-03-04..03-08·
summary
Argues that releasing differentially private synthetic microdata via a full cross‑classification with independent two‑sided geometric mechanism noise yields unbiased totals and means, and that total variance cleanly decomposes into sampling variance plus privacy noise. Presents simulations ( 5,000 replications, epsilon=1 ) showing near‑nominal 95% CI coverage and a roughly 10% increase in variance from noise. Proposes calibrating noisy table margins to fix negative cells, producing weighted microdata (which can create negative calibration weights) and keeps variance estimation straightforward.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:22Differential Privacy from a Statistical Perspective
- 6:25Official Statistics
- 11:36Geometric Mechanism
- 14:10Means and Totals
- 15:31Mean
- 16:46Example
- 21:56Sampling Uncertainty
- 23:03TwoStage Process
- 24:52TwoStage Process Example
- 29:48TwoStage Estimates
- 30:22Simulation Study
- 35:10Negative Cell Counts
- 41:13Illustration
- 42:20Conclusion
related talks
01h 03mStatistical Inference and Privacy, Part II
42mDifferential Privacy
01h 07mStatistical Inference and Privacy, Part I
20mShare data without revealing personal information
01h 00mPANEL: Statistical Theory, Privacy and Data Analysis
01h 00mDifferential Privacy: Fundamentals to Forefront I
30mAccuracy First: Selecting a Differential Privacy Level for Accuracy-Constrained...
31mWhen Existing Techniques in Linear Regression Preserve Differential Privacy