Communication Efficient Federated Learning via Channel-wise Dynamic Pruning

Federated Learning (FL) received widespread attention in 5G mobile edge networks (MENs) as it enables collaborative training deep learning models without disclosing users' private data. As the increasing number of parameters in the machine learning model poses a tremendous challenge for resource-constrained devices, there is a growing interest in applying model compression methods in federated learning. However, most existing model compression methods require a cumbersome procedure that introduces many additional hyperparameters and much more training time. In this paper, we propose a novel Channel-wise Dynamic Pruning method for communication efficient Federated Learning (FedCDP). The scheme dynamically evaluates the channel-wise parameter importance via a fast Taylor series evaluation and only communicates the important parameters in Federated Learning. Extensive experiments show the proposed method achieves both communication efficiency and model effectiveness in the benchmark datasets. The source codes are available at https://github.com/tabo0/FedCDP.