The Effect of Data Distribution on Federated Learning Stability: Experimental Insights with FedAvg and FedProx

Abstract

Data privacy and security have emerged as critical concerns in artificial intelligence applications. In this context, federated learning has gained attention as a method that prevents data from being transferred to a central server by enabling model training on local devices. In this study, the FedAvg and FedProx algorithms were employed to train models on the MNIST and FashionMNIST datasets. The impact of data distribution on model performance was examined by comparing Independent and Identically Distributed (IID) and nonIndependent and Identically Distributed (non-IID) scenarios. Experimental results showed that both algorithms achieved over 90% accuracy on FashionMNIST under IID settings, but performance decreased significantly in non-IID settings. Moreover, FedProx demonstrated slightly greater stability and robustness to data heterogeneity than FedAvg. These findings highlight the importance of considering data distribution and algorithm selection in the design of federated learningbased systems and underscore the need for simulations that better reflect real-world conditions. © 2025 IEEE.