Securing Federated GANs: Enabling Synthetic Data Generation for Health Registry Consortiums

Abstract

In this work, we review the architecture design of existing federated General Adversarial Networks (GAN) solutions and highlight the security and trust-related weaknesses in the existing designs. We then describe how these weaknesses make existing designs unsuitable for the requirements needed for a consortium of health registries working towards generating synthetic data sets for research purposes. Moreover, we propose how these weaknesses can be addressed with our novel architecture solution. Our architecture solution combines several building blocks to generate synthetic data in a decentralised setting. Federated GANs, Consortium blockchains, and Shamir Secret Sharing algorithm are the core building blocks of our proposed architecture solution. Finally, we discuss our proposed solution's advantages, disadvantages and future research directions.