EPFFL: Enhancing Privacy and Fairness in Federated Learning for Distributed E-Healthcare Data Sharing Services
Jingwei Liu, Yating Li, Mengjiao Zhao, Lei Liu, Neeraj Kumar · IEEE Transactions on Dependable and Secure Computing · 2024
Federated Learning (FL) has made remarkable achievements in medical and e-healthcare services. Different healthcare institutions can jointly train models to facilitate intelligent diagnosis. However, the model gradients transmitted among these institutions may still leak private information about the local models and training datasets. Additionally, in the current FL schemes, institutions with different quantities or qualities of medical data usually get the same training models, which may significantly hamper their motivation. Therefore, ensuring privacy and fairness in collaborative training remains a challenge. To address this issue, we propose a privacy-enhanced and fair FL scheme (EPFFL) to support distributed large-scale data sharing of e-healthcare services. In the training process, participants upload the encrypted model gradients according to their sharing wishes to the blockchain while storing their training data locally. Hence, the FL initiator can only get the aggregated gradients from the blockchain rather than the local data of other participants. Moreover, EPFFL ensures fairness by evaluating the participants’ contributions, i.e., participants with different data qualities and sharing levels can obtain the final models with different accuracies at the end of the training. Through theoretical and simulation analysis, the scheme shows superior functionalities on privacy preservation and fairness with the ideal model accuracy.