Mitigating Statistical Bias within Differentially Private Synthetic Data
This work addresses utility loss in privacy-preserving machine learning for applications relying on synthetic data, though it is incremental as it builds on existing differentially private generative models.
The paper tackles the problem of statistical bias in differentially private synthetic data, which reduces utility for downstream tasks, and proposes re-weighting strategies using privatised likelihood ratios to mitigate this bias, showing through large-scale evaluation that private importance weighting provides effective privacy-compliant augmentation.
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.