Cloud-Powered Federated Learning for Global Healthcare Diagnostics: Privacy Preserving Multi-Cloud Architecture

Abstract

Healthcare systems produce vast amounts of sensitive information from imaging, clinical notes, and electronic health records. While these data are invaluable for developing accurate diagnostic models, privacy regulations prevent direct sharing across institutions. Federated Learning (FL) provides a solution by enabling collaborative training without exposing raw data, but practical adoption in healthcare faces persistent challenges including non-IID data, communication overhead, and dependency on single cloud providers. This paper introduces a Multi-Cloud Federated Learning (MC-FL) framework that addresses these issues through four key innovations: (i) multi-cloud orchestration for compliance and fault tolerance, (ii) consent-driven filtering to enforce patient-level privacy policies, (iii) heterogeneity-aware adaptive aggregation, and (iv) two-stage model update compression for communication efficiency. Experiments on two benchmark healthcare datasets (NIH ChestX-ray14 and MIMIC-III) demonstrate that MC-FL achieves 98.21% accuracy, reduces membership inference attack leakage to 11.3%, and lowers communication cost by more than 50% compared to single-cloud FL. The framework also converges 15.8% faster while maintaining robustness as the number of clients scales from 5 to 20. These results show that MC-FL provides a practical and regulation-compliant approach for secure, scalable, and high-performance healthcare diagnostics across borders.