Blockchain-based Federated Learning with Contribution-Weighted Aggregation for Medical Data Modeling

To promote the sharing of medical data and thus to improve the medical level and resolve the imbalance of medical resource distribution is one of the most significant things of nowadays society. In this paper, a framework combining the newly emerging techniques, i.e., federated learning and blockchain, is proposed for decentralized medical data modeling and sharing. Federated learning is leveraged for data training and modeling, which will bring more precise medical models but without the leakage of original medical data. It can well solve the privacy concerns from patients. As the underlay for federated learning, blockchain contributes to provide a decentralized, secure, and transparent learning and sharing environment. In particular, a contribution-weighted incentive mechanism is proposed to promote the participation of medical data sharing, where the contributions and corresponding rewards are well considered and guaranteed. Finally, the prototype has been developed and implemented with a public data set of breast images. The results show that the proposed blockchain-enabled federated learning with contribution-weighted aggregation has advantages over the centralized learning approach and federated learning average aggregation in terms of model accuracy and system security.