Adaptive Asynchronous Federated Learning for Edge Intelligence

Edge intelligence has received great attention for the rapid development and wild application of edge computing and artificial intelligence. As a key technology in edge intelligence for its ability of privacy protection, federated learning faces many problems when deployed in an edge environment. The staleness effect causes by device heterogeneity make the performance of synchronous federated learning limited by these slower devices. And algorithm may also obtain a worse model for statistical heterogeneity. In this paper, we develop an adaptive asynchronous federated learning for edge intelligence to solve the above problems. To improve accuracy and stability of the algorithm, we realize a balance between synchronous and asynchronous settings and implement an adaptive optimization method. At last, we verify the performance in an experiment with heterogeneous settings.