Personalised Federated Learning in Healthcare: Balancing Accuracy, Fairness and Security

Abstract

Federated Learning (FL) evolved as an emerging approach for privacy-safe-guarding coordinated model training in healthcare, where raw patient data cannot be centralized due to regulatory and ethical constraints. However, the natural non-independent and identically distributed (non-IID) essence of clinical datasets across institutions degrades the performance of traditional FL algorithms such as FedAvg. To address this challenge, personalized federated learning (pFL) approaches have been developed, enabling client-specific adaptation while leveraging global knowledge. This paper provides a comparative analysis of three cutting-edge personalization algorithms, Per-FedAvg, pFedMe, and Ditto. Evaluated on healthcare datasets, including MIMIC-III and CheXpert. Beyond personalization, the study integrates essential security mechanisms, including Differential Privacy with Stochastic Gradient Descent (DP-SGD), Secure Aggregation (SecAgg) and Byzantine-robust aggregation, to defend against gradient inversion, membership inference and backdoor poisoning attacks. Experimental results demonstrate that pFedMe achieves the highest predictive accuracy (AUROC), Ditto ensures the most equitable performance across clients and Per-FedAvg enables rapid adaptation but remains sensitive to hyperparameters. The findings highlight a fundamental trade-off between fairness, convergence speed, and computational overhead in pFL. This work underscores the importance of designing FL frameworks that jointly optimize accuracy, security and fairness, paving the way for scalable and trustworthy deployment of federated learning in present healthcare systems.

Ähnliche Arbeiten