Private Data Synthesis from Decentralized Non-IID Data

Privacy-preserving data sharing enables a wide range of exploratory and secondary data analyses while protecting the privacy of individuals in the datasets. Recent advancements in machine learning, specifically generative adversarial networks (GANs), have shown great promise for synthesizing realistic datasets. In this work, we investigate the feasibility of training GAN models privately in practical settings, where the input data is distributed across multiple parties, and local data may be highly skewed, i.e., non-IID. We examine centralized private GAN solutions applied at each local party and propose a federated solution that provides strong privacy and is suitable for non-IID data. We conduct extensive empirical analysis with a wide range of non-IID settings and data from different domains. We provide in-depth discussions about the utility of the synthetic data, the privacy risks in terms of membership inference attacks, as well as the privacy-utility trade-off for private solutions.