pMSE Mechanism: Differentially Private Synthetic Data with Maximal Distributional Similarity
Talkat Data Privacy: From Foundations to Applications (Simons Institute workshop, 2019) · 2019-03-04..03-08·
summary
Introduces a mechanism that uses the exponential mechanism to sample model parameters and then generates differentially private synthetic datasets, optimizing a pMSE utility (propensity mean squared error) that measures distributional similarity. Works with arbitrary parametric synthesizers and unbounded continuous data; theory gives a 1/N sensitivity bound for globally optimal CART classifiers and a weaker 0.85/m partition bound in practice. Simulations show the pMSE mechanism outperforms smooth histograms at low epsilons, violations of the strict bound occur <5% empirically, and practical limits include MCMC sampling issues and high-dimensional parameter spaces.
AI-written summary · about
chapters (19)
- 0:00Intro
- 0:50Privacy Conversations at a Policy Institution
- 2:05What are the Drivers of This Disconnect?
- 5:31Background on Synthetic Data
- 6:27Common (Non-Differentially Private) Synthesis Methods
- 7:16Purpose of Synthetic Data in Contrast to Queries
- 8:18How Do We Evaluate Synthetic Data?
- 17:17Resampling Techniques
- 17:48Motivation for Differentially Private Synthetic Data
- 19:01Motivating the Need for Privacy Constraints
- 20:34Common Approaches to DP Synthetic Microdata
- 21:37Accuracy Depends on the Assumed Bound of Continuous Data
- 23:59Making the Synthesis Process DP
- 24:41Choosing the Utility Function
- 26:06Bounding the Sensitivity (Snoke and Slavković 2018)
- 26:40Summarizing the PMSE Mechanism
- 28:02Optimal Trees vs. Greedy Fitting
- 28:39Empirical Sensitivity Simulations
- 31:16Simulation Results for Evaluated PMSE on DP Synthetic Data
related talks
13mDPSynth: From Research to Production—Engineering Differentially Private Synthetic...
14mEfficient Universal Estimators For Symmetric Property Estimation
01h 14mStatistical Inference and Privacy, Part III
01h 02mLarge-Scale Private Learning, Part I
48mRobust Statistics and Bias Correction: Ideas for Differential Privacy and Utility
39mPufferfish Privacy Mechanisms for Correlated Data
44mPrivate Meta-Selection
44mDifferentially Private Inference for Regression Coefficients