Information Security and Federated Learning for Collaborative Healthcare Diagnosis
TL;DR
A federated learning framework with data locality for healthcare diagnostics that enables multiple hospitals to jointly train accurate models while keeping patient data within institutional boundaries is introduced.
Abstract
Healthcare institutions produce enormous quantities of medical data daily. However, regulations such as HIPAA, GDPR, and DPDPA create strict barriers to sharing patient information for collaborative artificial intelligence development. This forces individual hospitals to develop models using only their local data, producing systems that fail to generalize across diverse populations. We introduce a federated learning framework with data locality for healthcare diagnostics that enables multiple hospitals to jointly train accurate models while keeping patient data within institutional boundaries. From an information security perspective, the system follows a coordinator-client architecture using the Flower framework with a FastAPI backend and a Next.js frontend, facilitating distributed training under data locality. Medical imaging remains entirely within local hospital infrastructure while encrypted parameter updates traverse secure channels. The system implements Federated Averaging, Federated Optimization, Federated Matched Averaging, and our novel Performance-Adaptive Weighted Aggregation algorithm. Our approach demonstrates the feasibility of collaborative training under data locality in limited experiments; a direct comparison with centralized training was not performed.
