Practical Experience with Making Synthetic Data Differentially Private
Talkat Data Privacy: From Foundations to Applications (Simons Institute workshop, 2019) · 2019-03-04..03-08·
summary
Practical work tests three concrete ways to make synthetic tabular data differentially private: Laplace on full table, noisy marginals + IPF, and MWEM (multiplicative weights exponential mechanism), applied to a 5,000‑row Poland survey and a 75,000‑row 1911 Scottish census extract with seven categorical variables (~12k cells). Method one only preserves utility when signal far exceeds noise (very large ε); method two gives better utility but needs fiddly constrained marginals and a correct generative model; MWEM looks promising but is unstable and requires choosing T. Utility uses synth‑pop PMSE/chi‑square; DP reduces simple attribute disclosure but results are data‑ and parameter‑sensitive.
AI-written summary · about
chapters (14)
- 0:00Start
- 6:55Synthetic extracts are not open
- 8:41Overview
- 11:00Correct generative model assumption
- 12:51Models for data synthesis
- 15:16Synthesis by iterative proportional fitting
- 21:46DP syntheses
- 22:23DP synthesis method 1
- 32:43DP method 2 a hybrid method
- 34:19DP method 3 MWEM Multiplicative Weights Exponential Mechanism
- 37:40Method 3 summary
- 38:00A disclosure measure for synthetic data
- 38:43Correct Relative Attribution Probability
- 42:42Final thoughts
related talks
01h 14mStatistical Inference and Privacy, Part III
44mDifferentially Private Inference for Regression Coefficients
13mDPSynth: From Research to Production—Engineering Differentially Private Synthetic...
20mShare data without revealing personal information
16mGenerating High-Quality Tabular Synthetic Data at Scale
30mAccuracy First: Selecting a Differential Privacy Level for Accuracy-Constrained...
01h 00mPANEL: Statistical Theory, Privacy and Data Analysis
19mPractical Considerations for Differential Privacy