Privacy-Preserving Federated Generative Adversarial Network for IoT

Federated Learning (FL) shows its great vitality for terminal devices of Internet of Things (IoT) in privacy protection. However, the amount of data on each device in IoT are imbalanced which makes the global training model rarely effective. Although Generative Adversarial Network (GAN) can generate data to alleviate the above problem, it has the characteristic of training instability and still carries the risk of privacy breach. In this paper, we propose a novel privacy-preserving federated GAN framework, named P-FedGAN, to train a federated generative model with privacy protection function. Each device uses the trained model to generate enough synthetic data to replace sensitive data to train desired model that can complete typical data analysis tasks. We carry out several sets of experiments to test the validity of proposed framework. The results of the experiments indicate that proposed framework meets differential privacy constraints and produces high-quality model utility compared to federated average algorithm simultaneously.