Fedhealth-Privacy: Private Federated Learning for Multi-Hospital Medical Imaging

Medical imaging data is highly sensitive and naturally siloed across hospitals and research institutions, severely limiting the training of robust deep learning models for disease detection. Federated Learning (FL) enables collaborative model training without raw data exchange, but remains vulnerable to gradient leakage attacks and membership inference. We propose FedHealth-Privacy, a novel FL framework that integrates differential privacy (DP) with adaptive gradient clipping and secure multi-party aggregation. Evaluated on three real-world chest X-ray datasets (NIH ChestX-ray14, CheXpert, MIMIC-CXR) under realistic non-IID splits (prevalence and acquisition shift), our method achieves 95.2% AUROC for pneumonia detection—comparable to centralized training (96.1%)—while providing a provable privacy budget (ε = 3.5, δ = 1e-5). We further demonstrate robustness against gradient inversion attacks (reconstructed image SSIM < 0.05) and membership inference (advantage < 0.02). FedHealth-Privacy offers a deployable, privacy-compliant solution for multi-institutional medical AI.