Title: Practical tools for privacy preserving synthetic data generation 

Abstract: Synthetic data generation is quickly becoming a promising way to share sensitive data without violating data privacy. However, differential privacy (DP) based generative modeling strategies often lead to very poor sample quality, which constrains their downstream use. In the first part of my talk I will present privGAN, a novel generative machine learning framework which enables the generation of privacy-preserving synthetic data without sacrificing much downstream utility. PrivGAN focuses on protecting against a narrower notion of privacy than DP, namely the membership privacy risk, which is motivated by real-world privacy concerns in healthcare data sharing. This in turn leads to the generation of synthetic data which has a much better privacy/utility trade-off than DP based approaches. 

In the second part of my talk I will present a novel membership privacy auditing framework for generative models, called MACE. While there has been much work on non-DP generative modeling since privGAN, there currently exists no formal way to quantify the privacy loss of such models. MACE poses membership inference as a classification problem and shows how Bayes optimal classifiers can be used to estimate the expected membership privacy loss of models for a given dataset. In addition to estimating membership privacy loss of models, our framework is also able to identify the membership privacy risk to individual samples.