A Robust Federated Learning Framework for Healthcare Across IID and Non-IID Data Distributions

Abstract

Privacy-preserving machine learning has become essential in healthcare, where sensitive patient data cannot be centralized without risking confidentiality and regulatory non-compliance. Federated learning (FL) offers a viable alternative by enabling collaborative model training while retaining data on local medical institutions. This study presents a robust federated learning framework designed to maintain strong predictive performance across both independent and identically distributed (IID) and non-IID data scenarios, reflecting realistic variability in healthcare environments. Using EfficientNet-B0 as the core architecture and the PathMNIST dataset as the benchmark, we evaluate the framework across federations of 5 and 10 clients, systematically comparing centralized and federated setups. Experimental results demonstrate that the proposed framework achieves 95.29% accuracy under IID conditions with 5 clients and 94.99% with 10 clients, closely matching centralized performance. Under non-IID distributions generated via a Dirichlet partitioning, the framework maintains competitive performance with 94.26% accuracy for 5 clients and 93.20% for 10 clients. Additional metrics highlight the system’s robustness: precision reaches up to 95.38%, recall up to 95.35%, and F1-score up to 95.23% in centralized benchmarking, with only marginal degradation under federated settings. Convergence curves show stable optimization in IID scenarios and controlled fluctuations under non-IID heterogeneity, confirming the resilience of the federated averaging strategy. These findings demonstrate that the proposed federated learning framework delivers high model utility while ensuring decentralized data governance, making it suitable for scalable, privacy-conscious medical image analysis.