Convergence Analysis of Cloud-Aided Federated Edge Learning on Non-IID Data

Federated edge learning has attracted great attention for edge intelligent networks. Due to the limited computation and energy, mobile devices usually need to offload data to nearby edge servers. Facing this scenario, we design a cloud-aided federated edge learning (CA-FEEL) framework where the edges cooperate with the cloud to train a federated learning model. Specifically, the edges adopt the gradient descent (GD) method in parallel to update the edge parameters and the cloud averages them to update the global parameter. By theoretical analysis, we find that the covariance of non-independent and identically distributed (non-IID) data sets hinders the convergence of the GD based FL. Thus, we propose a CA-FEEL algorithm by adding a simple judgment condition. It is proved to have a theoretical guarantee of convergence for convex and smooth problems. Experiment results indicate that the proposed algorithm outperforms the standard federated learning in terms of the convergence rate and accuracy.