Efficient federated learning with adaptive channel pruning for edge devices

Federated learning (FL) is an emerging machine learning paradigm that allows distributed participants to train a global model collaboratively. In edge networks, FL involves numerous edge devices updating the local models with the assistance of a central server while keeping the local data on the devices. However, edge devices are commonly resource-limited while neural network models used in FL require significant resource consumption to reach a satisfactory accuracy. Moreover, different devices are various in terms of system capabilities, and employing a uniform model on all devices leads to the derogation of model accuracy. Several pioneer work integrate model pruning techniques into the FL process and focus on reducing the resource consumption on edge devices, whereas neglecting the impact of model generality on accuracy. In our work, we propose FedACP, an efficient FL approach with adaptive channel pruning for edge devices to achieve better resource-accuracy trade-offs. Specifically, we design a two-phase channel pruning approach to adaptively prune local models, in which both the resource constraints and model generalities are taken into consideration. In addition, we also design an aggregation algorithm for aggregating heterogeneous local models produced by the two-phase channel pruning approach. To evaluate the performance of the FedACP, we implement it with a real FL framework FedML and compare it with several state-of-the-art methods. The experimental results show that FedACP achieves better model accuracy while simultaneously reducing 22.41% to 27.34% model parameters.