Privacy-Preserving Federated Learning for Secure Healthcare Data: A Sensitivity-Aware Selective Encryption Framework for Cross-Institutional Clinical Model Training

Federated learning (FL) lets hospitals train a shared model without pooling patient records, but exchanged updates can still leak sensitive information, and homomorphic encryption (HE) of every parameter is costly for resource-limited clinical sites. This paper proposes SASHE-FL, a sensitivity-aware selective HE framework for cross-silo healthcare FL. Sites use SCAFFOLD-style local training and encrypt, with the CKKS scheme, a jointly agreed top-p fraction of parameters ranked by a label-sensitivity score; the server aggregates ciphertexts without decrypting them. A prototype with real CKKS encryption (TenSEAL) was evaluated on the Wisconsin breast-cancer dataset, split non-IID across five simulated institutions. Accuracy matched plaintext SCAFFOLD at every encrypted fraction (96.5 ± 1.4%). Encrypting 10% of parameters cut per-site upload by 81.5% and cryptographic compute time by 86.3% versus encrypting all parameters. However, a worst-case gradient-inversion probe still recovered inputs (cosine similarity 0.999) with the top 30% encrypted, and failed only when the entire first layer was encrypted. SCAFFOLD gave no accuracy gain over FedAvg. Protection covers only encrypted coordinates, against a non-colluding semi-honest server. The results are a small proof of concept, not clinical evidence or a formal privacy guarantee.