Less is More: Variance-Gated Client Rejection for Faster Convergence in Federated Learning

Abstract

This paper proposes VGCR-FL, a method that speeds up federated learning by filtering out client updates that deviate strongly from the rest of the cohort. The filter uses a threshold that adapts each round, plus a check on whether the update points in the same direction as the cohort. This keeps useful updates from minority clients while cutting the variance and communication cost of aggregation. The paper sets out the method, a convergence analysis, a fairness-aware extension, and a reproducible evaluation plan (CIFAR-10, FEMNIST, Shakespeare). No experimental results are reported yet.