Optimizing Federated Learning on Non-IID Data with Clustering and Model Sharing

The increased amount of data generated by edge computing has necessitated the development of efficient methods to leverage this vast information. Federated Learning (FL) offers a promising solution by enabling distributed model training while preserving privacy. However, FL faces challenges with NonIndependent and Identically Distributed (Non-IID) data, which can impact model accuracy and convergence. To address this, we propose a three-stage framework that effectively trains models in Non-IID scenarios. Our approach effectively reduces the effect of Non-IID by classifying IID clients, evaluating model performance through model sharing, and dynamically adjusting the weighting of each client to perform client selection. We created extreme Non-IID environments with diverse client representation. Tested on the CIFAR-10 dataset, our method improves maximum accuracy by up to 3.03 % compared to other state-of-the-art and traditional methods such as FedProx and FedAvg, demonstrating its effectiveness in Non-IID scenarios.