Robust Federated Learning for Heterogeneous Clients and Unreliable Communications

Federated Learning (FL) serves as a machine learning paradigm where distributed devices collaboratively train on local data, with their models subsequently aggregated on a central server. However, challenges arise due to unreliable communication channels, potential sign errors in model parameters, data heterogeneity, and resource limitations that can hinder full client participation. In this paper, firstly, we address these issues by proposing an optimization objective that minimizes FL loss while taking into account constraints on delay and energy consumption. Secondly, to counteract the model drift caused by data heterogeneity and packet errors, we introduce a proximal term in the local training process and incorporate packet errors into the global aggregation phase. Finally, we establish a theoretical convergence upper bound for our FL algorithm in complex non-convex situations, providing insights to guide the formulation of client sampling strategies and ensure FL algorithm convergence. We validate our algorithm’s superior accuracy on the MNIST and CIFAR-10 datasets.