A Secure and Scalable Federated Deep Learning Framework for Distributed Medical Imaging

Abstract

The digitization of healthcare has yielded unprecedented volumes of heterogeneous medical imaging data distributed across geographically dispersed hospitals, diagnostic centers, and clinical research institutions. Conventional centralized deep learning approaches require raw patient data to be aggregated on a single server, creating critical tensions with patient privacy legislation—including the Health Insurance Portability and Accountability Act (HIPAA) and the European General Data Protection Regulation (GDPR)—and introducing single-point-of-failure vulnerabilities in mission-critical clinical infrastructure. Federated Learning (FL) addresses these concerns by enabling collaborative model training across distributed data silos without raw data exchange; however, existing FL systems for medical imaging remain constrained by gradient-inversion privacy vulnerabilities, Byzantine-robust aggregation deficiencies, communication bandwidth bottlenecks, and statistical heterogeneity arising from non-IID imaging distributions. To address these challenges comprehensively, this paper proposes SecFedMed—a Secure and Scalable Federated Deep Learning Framework for Distributed Medical Imaging—integrating four tightly coupled innovations: (1) a Rényi Differential Privacy (RDP) mechanism with adaptive noise calibration for formal gradient-level privacy guarantees; (2) a Secure Aggregation Protocol (SAP) employing threshold secret sharing and homomorphic encryption; (3) an Adaptive Model Compression (AMC) module reducing per-round communication cost by 94.7%; and (4) a Heterogeneity-Aware Federated Proximal Aggregation (HAFPA) strategy counteracting non-IID performance degradation.