Dynamic Logical Resource Reconstruction against Straggler Problem in Edge Federated Learning

Federated learning enables edge distributed devices to learn a shared global model without violating privacy concerns. However, due to the heterogeneity in data or resources across devices, the straggler issue has become a key bottleneck for effective federated learning. As aggregation methods favor faster devices, these methods may introduce biases and severely degrade model accuracy. To solve these challenges, we propose FedDRB, a unique federated learning communication framework via dynamic resource balancing, including a dynamic logical cluster construction (DLCC) algorithm and a weighted intra-cluster collaborative (WICC) aggregation algorithm. To shorten total model training time, DLCC divides all edge devices into several logical clusters and constructs a tiered structure. In addition, WICC requires faster devices to assist the training of slower devices and assigns a relatively higher weight to slower devices during the aggregation stage, hence accelerating the intra-cluster convergence speed and ensuring training model accuracy. Compared with stateof-the-art federated learning approaches, experimental results demonstrate that FedDRB improves prediction accuracy by up to 12.60% and reduces the required communication cost by up to 9.78x.