Asynchronous Byzantine-Resilient Distributed Optimization with Momentum
Yi Wan, Yifei Qu, Zuyan Zhao, Shaofu Yang · 2022 41st Chinese Control Conference (CCC) · 2022
In this paper, we propose a novel asynchronous Byzantine-resilient distributed optimization algorithm. In our algorithm, we adopt multiple buffers in the server to regulate the asynchronous update of workers. In addition, to cope with the Byzantine workers, we combine robust aggregate operator with momentum-based update, which utilize the historical information to improve the resilience of distributed algorithm. It is shown that momentum is beneficial for resisting time-coupled attacks. We theoretically prove the convergence of our algorithm at a speed of$O(1/\sqrt{T})$in the nonconvex setting. Finally, simulation results are provided to show the effectiveness of our algorithm.