A loss-based weighted aggregation method for federated learning with heterogeneous computing resources

Federated Learning (FL) faces challenges in model convergence and stability due to the heterogeneity of client resources, where different local batch sizes degrade global model performance. We propose FedLW, a dynamic client weighting approach that utilizes training loss values to assess model quality and assigns higher weights to clients with lower losses. This way enables the server to prioritize gradients from well-performing clients, take full advantage of different clients at different training stages, and improve the generalization ability of the global model. To validate the effectiveness of the method, experiments on MNIST, Fashion-MNIST, and CIFAR-10 show that FedLW improves the accuracy by up to 2.24% compared to standard FL aggregation methods.