Adaptive Training-Communication-Aggregation for Heterogeneous Federated Learning

Abstract

Federated learning (FL) often suffers from reduced training efficiency and degraded global model performance due to system heterogeneity in computation and communication, as well as data heterogeneity. We present an adaptive framework that integrates three complementary mechanisms across training, communication, and aggregation to mitigate these issues. First, an adaptive local-round scheduler (ALRS) adjusts the number of local epochs based on device processing time, providing a simple yet effective means to control computational imbalance among straggling and fast devices, and reduce cross-device idle waiting. Second, a time-aligned bandwidth allocation (TABA) algorithm improves communication synchronization by redistributing uplink physical resource blocks (PRBs) based on the remaining communication time gap between the most and least delayed devices. Third, an entropy-based hybrid aggregation (EHRCA) mechanism limits the influence of highly skewed or overly dominant devices through entropy-guided risk scoring, direction-aware update correction, and adaptive reweighting, thereby reducing bias from skewed updates while preserving the contributions of well-balanced devices. Simulation results indicate that, under diverse system and data heterogeneity, these mechanisms better align computation and communication times across devices, leading to more synchronized aggregation rounds and improved training stability and accuracy.