FedHealthcare: Federated Learning and Lightweight Additive Homomorphic Encryption‐Based Privacy‐Preserving Healthcare

ABSTRACT In recent days, collaborative health data analysis is conducted in various health organizations. Hence, data privacy and security are major concerns for healthcare industries. The existence of strict regulations underscores the urgent need for secure and compliant data‐sharing solutions. To that aim, this paper proposes FedHealthcare, a privacy‐preserving machine learning (ML) framework that integrates federated learning (FL) with lightweight additive homomorphic encryption (HE). This scheme allows every healthcare organization to train a local model, and it uses lightweight additive HE to encrypt the sensitive parameters. After every round, all clients receive the encrypted updates that have been safely combined on a global server via homomorphic addition. This conceals the raw data. Compressed gradient aggregation and adaptive encryption preserve high accuracy and privacy rules while consuming less bandwidth and computation. Not only does it encrypt the sensitive model parameters, but it also integrates the compressed gradient aggregation. This improves training efficiency without compromising accuracy. Experiments are conducted on realistic healthcare datasets. An accuracy achievement of more than 90.8% is possible using FedHealthcare with lower bandwidth usage (250 KB/round) and a 20% improvement in encryption speed compared to full HE approaches. The improved results demonstrate that the integration of FL and HE works well to protect privacy while preserving high model performance. This makes FedHealthcare a good option for extensive medical AI applications.