A differential privacy federated learning framework for accelerating convergence

Federated Learning (FL) is a special distributed machine learning environment. It is jointly trained by many clients under the coordination of a central server. And differential privacy can provide privacy guarantee for FL. While, federated learning, compared with centralized learning, converges at slower speed. And differential privacy exacerbates this trend. In this paper, we propose a novel FL algorithm named DP-FedADMM to solve these problems. We merge differential privacy (DP) into the FedADMM algorithm and propose a method to handle noisy gradients. Under the guidance of the help of the latest results in differential privacy theory, we provide a privacy proof of DP-FedADMM. Through extensive experiments on datasets, it is demonstrated that DP-FedADMM outperforms the currently popular DP-FedAvg algorithm in terms of model convergence speed.