A GAN-based federated learning architecture for data augmentation of medical images

Abstract

n the medical industry, the availability of precise data limits the scope of deep learn ing applications. Institutional norms restrict hospitals and research facilities owing to privacy concerns. Therefore, data collection from such sources is unfeasible. Fed erated Learning (FL) is promising in this scenario, but it does not guarantee data privacy. In this paper, we will use Deep Convolutional Generative Adversarial Net work (DCGAN) and Wasserstein Generative Adversarial Network (WGAN) on an OCT dataset to demonstrate that the Federated GAN (FedGAN) architecture fails in these networks due to its innate structure. Additionally, introduce a Distributed Generative Adversarial Network (Distributed GAN) that collects and distributes the weights of each temporary GANs on the client side to the main server to tackle the mode collapse risk of non-iid data. This conserves the optimal distribution of data to all private discriminators while protecting sensitive individual data.