Solving Non-IID in Federated Learning for Image Classification using GANs
Thiti Chuenbubpha, Thapana Boonchoo, Jason Haga, Prapaporn Rattanatamrong · 2023
Federated Learning (FL) has emerged as a powerful methodology for training centralized models while preserving data privacy by using trained parameters from local models that are distributed among decentralized sites. Despite its growing popularity in the development of cloud-based Internet of Things (IoT) applications, FL performance is significantly impacted when data is non-independently and identically distributed (non-IID). This paper proposes a novel framework, called GANs Augmented IID - Federated Learning (GAIID-FL) to tackle the diversity of data distribution among clients in FL. The GAIID-FL framework collaboratively trains Generative Adversarial Networks (GANs) and then employs the trained GANs models to generate synthetic data and distribute proportionally to each device. Using GAIID-FL effectively restores the IID data distribution for the setting. The experimental results demonstrate that our framework can achieve up to 45% improvement in accuracy. In addition, the batch collaborative training approach for GANs models can reduce communication overhead by up to 90 times when compared to the unoptimized method.