Federated diabetes mellitus analysis via homomorphic encryption

Abstract

Diabetes is one of the most extensive chronic diseases in the world. Most patients suffer from the disease and its complications for a long time due to the lack of accurate and standardized treatment at early stage. Therefore, by analyzing diabetes data and establishing relevant predictive models, it is very meaningful to give reasonable health advice to high-risk groups. The establishment of an accurate prediction model requires a large number of data sources as support, and multiple medical institutions participate in data contribution and collaborative learning. Federated learning provides a secure general architecture for distributed collaborative learning. However, in the process of federated learning, the data of participants is still subject to the risk of security attacks or indirect information leakage. For example, when a participant uploads the local model parameters to an honest but curious cloud server, the cloud server can obtain relevant information from the participant. In order to solve this problem, this paper proposes a federated forest algorithm based on homomorphic encryption to strengthen the protection of the data privacy of the participants while ensuring that the accuracy of data analysis does not decrease. Analysis proves that our algorithm has good performance in privacy protection and prediction accuracy.