FedDP: Secure Federated Learning for Disease Prediction with Imbalanced Genetic Data

Abstract

It is challenging to share and aggregate biomedical data distributed among multiple institutions or computing resources due to various concerns including data privacy, security, and confidentiality. The federated Learning (FL) schema can effectively enable multiple institutions jointly perform machine learning by training a robust model with local data to satisfy the requirement of user privacy protection as well as data security. However, conventional FL methods are exposed to the risk of gradient leakage and cannot be directly applied to genetic data since they cannot address the unique challenges of data imbalance typically seen in genomics. To provide secure and efficient disease prediction based on genetic data distributed across multiple parties, we propose an FL framework enhanced with differential privacy (FedDP) on trained model parameters. In FedDP, local models can be trained among multiple local-hold genetic data with efficient secure and privacy-preserving techniques. The key idea of FedDP is to deploy differential privacy on compressed intermediate gradients that are computed and transmitted by optimizers from local parties. In addition, the unique weighted minmax loss in FedDP is able to address the difficulties of prediction for highly imbalanced genetic datasets. Our experiments on multiple genetic datasets demonstrate that FedDP provides a powerful tool to implement and evaluate various strategies in support of privacy preservation and model performance guarantee to overcome data imbalance.