Differentially Private Inference for Regression Coefficients
summary
Creates a synthetic version of the Office of Personnel Management federal employee dataset and builds a web verification server that answers regression queries under differential privacy. Uses Laplace noise and a partitioning strategy (split data into k subsets) to bound sensitivity, then aggregates subset estimates to produce DP posterior distributions, point estimates, sign probabilities and confidence intervals for individual regression coefficients. UI lets users pick epsilon, number of partitions and submit coefficient queries to see DP intervals and posterior probabilities.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 5:00Alpha synthetic version design
- 10:00Data structure and models
- 15:00Verification methods and servers
- 20:00Private inference in regression
- 25:00Probability in models
- 30:00Practical application and results
- 35:00Using verification server
- 40:00Conclusiones y pasos futuros
- 42:30Q&A
related talks
45mPractical Experience with Making Synthetic Data Differentially Private
31mWhen Existing Techniques in Linear Regression Preserve Differential Privacy
27mYou've got trust issues, we've got solutions: Differential Privacy
26mAaron Roth, Associate Professor, University of Pennsylvania, NYC 2017
01h 07mStatistical Inference and Privacy, Part I
10mCompiling Python Programs into Differentially Private Ones
19mPractical Considerations for Differential Privacy
32mYou've got trust issues, we've got solutions: Differential Privacy