A Practical Method to Reduce Privacy Loss when Disclosing Statistics Based on Small Samples
Talkat Data Privacy: From Foundations to Applications (Simons Institute workshop, 2019) · 2019-03-04..03-08·
summary
Releases local regression predictions with added noise using a data-driven sensitivity cap called maximum observed sensitivity (MOS) (Chi scales as Chi/n), applied to a 20 million-person Census–tax panel; uses Windsorization at 5–95%, Gaussian noise, state×race×gender cells, and censors cells <20 per IRS. Method is not provably private but reduces bias versus count-suppression (example: teenage birth estimate rises from 34% to 40% across a covariate shift), retains usable precision for cell sizes ~50–200, and produces the public Opportunity Atlas maps now viewed by >500,000 users.
AI-written summary · about
chapters (15)
- 0:00Intro
- 2:00Publishing Statistics Based on Small Cells
- 2:41Intergenerational Mobility in the United States Mean Child Household Income Rank vs. Parent Household Income Rank
- 6:27Controlling Privacy Loss
- 6:47This Paper: A Practical Method to Reduce Privacy Loss
- 10:22Calculating Sensitivity
- 11:53Maximum Observed Sensitivity
- 16:06Data Sources and Sample Definitions
- 17:53Specification Details
- 23:40Other Practicalities for Privacy Method Implementation
- 26:18Application: Opportunity Atlas
- 33:06Comparison to Alternative Methods: Privacy
- 33:51Comparison to Alternative Methods: Statistical Bias
- 35:30Teenage Birth Rates for Black Women vs. Number of Black Women with Teenage Births in Tract
- 42:20Future Work
related talks
20mShare data without revealing personal information
48mRobust Statistics and Bias Correction: Ideas for Differential Privacy and Utility
01h 00mPractical Difficulties in the Application of Differential Privacy for the Release of Large-Scale...
17mNegotiating Privacy/Utility Trade-Offs under Differential Privacy
47mBlack-box Random Differential Privacy with Sensitivity Sampling
01h 10mThe Social Survey Statistician's Perspective
27mYou've got trust issues, we've got solutions: Differential Privacy
01h 03mStatistical Inference and Privacy, Part II