Mitigating Statistical Bias within Differentially Private Synthetic Data

TL;DR

This paper introduces re-weighting strategies using privatized likelihood ratios to reduce statistical bias in differentially private synthetic data, improving utility for downstream tasks.

Contribution

It proposes a novel private importance weighting method that mitigates bias and enhances the utility of synthetic data in privacy-preserving machine learning.

Findings

Private importance weighting effectively reduces bias in synthetic data.

The approach improves downstream predictive model performance.

Method shows broad applicability across different generative models.

Abstract

Increasing interest in privacy-preserving machine learning has led to new and evolved approaches for generating private synthetic data from undisclosed real data. However, mechanisms of privacy preservation can significantly reduce the utility of synthetic data, which in turn impacts downstream tasks such as learning predictive models or inference. We propose several re-weighting strategies using privatised likelihood ratios that not only mitigate statistical bias of downstream estimators but also have general applicability to differentially private generative models. Through large-scale empirical evaluation, we show that private importance weighting provides simple and effective privacy-compliant augmentation for general applications of synthetic data.