A Simple Data Privacy and Security Protection Algorithm Based on Federated Learning

Abstract

Abstract Although federated learning can break the data island and enhance privacy, it is not absolutely secure. External eavesdroppers can intercept the model when clients upload their models to the server, which might be used to infer the data information. A natural solution is to add noise to the uploaded model. However, this solution has its drawback, because the added noise cannot be eliminated when the models are aggregated at the server, which leads to poor performance. In order to tackle this problem, we propose a simple but effective algorithm, FedNoise. In the vanilla federated learning algorithm, a gradually decreasing learning rate is applied at the client side. While in our algorithm, we divide the parameter update into client-side and server-side. The stochastic gradient descent update with constant learning rate is performed at the client side and a gradually decreasing learning rate is performed at the server. Our algorithm can not only ensure the data security, but also maintain comparable system performance. Simulation results validate our conclusion.