A Robust Federated Learning with Bounded Gaussian Mechanism

Federated learning is a distributed machine learning method that allows multiple devices or servers to work together to train a shared model while maintaining data localization and privacy, and has attracted widespread attention. In the complex learning process, if some malicious participating devices or adversaries alter the upload parameters during communication, it will reduce the training accuracy. The approach with standard arithmetic mean aggregation cannot overcome this problem. This paper proposes a robust federated learning framework via a novel robust aggregation, called Huber aggregation, to avoid malicious attacks. Moreover, we propose a bounded Gaussian mechanism to enhance the privacy of the Huber aggregation based federated learning and analyze its basic property in privacy. We use a simulation example of the convex clustering to verify the effectiveness of the proposed method.