AdaptFL: Adaptive Client Task Allocation-Based Synchronous Federated Learning

Federated learning is a privacy protection method for machine learning, enabling the maintenance of data localization and privacy during training. However, the system heterogeneity leads to different time consumption for each client to complete each round of local training, thereby causing a serious straggler effect and affecting the overall efficiency of federated learning. Existing work has addressed the system heterogeneity issue by proposing task allocation strategies based on client capabilities to solve the straggler effect. However, simply adjusting the training tasks based solely on client capabilities can cause a decrease in the accuracy of local models, affecting the global model’s accuracy. In this paper, we propose AdaptFL, an adaptive client task allocation-based synchronous federated learning algorithm. This algorithm selects clients with a high probability of participation in training in each round. It adjusts the number of local epochs for clients dynamically based on training speed and the loss of local model. AdaptFL can alleviate the straggler effect effectively without reducing model accuracy, thereby enhancing the overall training efficiency of federated learning. Finally, we evaluate AdaptFL using the CIFAR-10 and FMNIST datasets, and the results demonstrate that the convergence and efficiency of AdaptFL.