A Healthcare Data Sharing Scheme Based on Homomorphic Encryption and Federated Metric Learning

Medical data is highly sensitive, often leading to “data silos” in sharing processes. To address this issue, this paper combines blockchain and federated metric learning technologies, and innovatively integrates them with homomorphic encryption technology to propose a privacy protection system suitable for application in medical data sharing scenarios. The system employs federated learning to train a global model across distributed client nodes without sharing raw data. Metric learning classifies clients as benign or malicious based on data similarity, only aggregating models from benign clients to improve efficiency and mitigate data distribution issues. Additionally, differential privacy and homomorphic encryption are used to further protect data privacy by preventing the original data from being reverse deduced and model parameter leakage during aggregation. Experimental results show that the system can achieve a prediction accuracy of 89.102564% under different data distributions, and its performance on medical datasets has been significantly improved compared to traditional federated learning.