FEDMI: A Federated Learning Framewoek for Secure Sharing of Medical Images

A large amount of medical image data about patients can be shared and applied to build a medical image AI system, which is conducive to rapid and accurate diagnosis and treatment of diseases. However, privacy leakage and data heterogeneity problems in the data-sharing process pose challenges to model construction. To address the above problems, a model construction method with privacy protection under cloud edge architecture is proposed. First, the edge server model is partitioned into two parts during the training process, the output representation of the first part of the model is fitted by generative adversarial network, and then the model parameters with the latter part are uploaded to the central cloud server after differential privacy processing to aggregate and generate the global model and send it down. The experimental analysis of the proposed scheme can realize medical data sharing while protecting the privacy of medical data.