Federated Learning with Differential Privacy for Cross-Silo Medical Image Analysis

The aggregation of medical imaging data across healthcare institutions offers significant potential for improving diagnostic AI systems, yet patient privacy regulations and data sovereignty concerns severely restrict data sharing. We present FedMed-DP, a federated learning framework that incorporates Rényi differential privacy to enable privacy-preserving collaborative training across hospital silos. Our approach introduces adaptive noise calibration based on layer sensitivity analysis, reducing privacy budget consumption by 34% compared to standard DP-SGD while maintaining model utility. Applied to chest X-ray pathology detection across 8 hospital datasets (124,442 images total), FedMed-DP achieves AUC of 0.943, matching centralized training performance within 2.1%, while providing (ε=1.2, δ=10⁻⁵)-differential privacy guarantees.