Abstract
Abstract
Machine learning is increasingly used in healthcare for risk prediction and clinical decision support, yet legal, ethical, and organisational constraints frequently prevent direct pooling of electronic health record (EHR) data. Federated learning (FL) addresses this limitation by retaining records within participating institutions. However, plaintext model updates may still disclose information to an honest-but-curious aggregation server or to an adversary observing the communication channel. This paper presents a CKKS-based homomorphic federated learning (HE-FL) framework for privacy-preserving prediction across multiple simulated hospitals. Ten clients train locally on non-identically distributed partitions of the Diabetes 130-US Hospitals dataset and transmit encrypted parameter updates for server-side aggregation. The framework is compared with centralised learning, plaintext FL, and differential-privacy baselines using predictive, privacy, and systems metrics. Plaintext FL achieved the highest utility (F1 = 0.9922; ROC-AUC = 0.9996), whereas the proposed HE-FL configuration retained meaningful predictive capability (F1 = 0.4042; ROC-AUC = 0.7714) and reduced membership-inference attack performance to the random-guess level (AUC = 0.5000) under the evaluated protocol. This privacy improvement incurred measurable systems cost, including approximately 24× ciphertext expansion and an uplink increase from 3.18 MB to 37.69 MB per round. We show that encrypted aggregation can help reduce the leak of the intermediate updates. However, in practice, the deployment needs to explicitly consider the communication overhead, convergence degradation, and utility loss of the model constraints compatible with encryption.