Communication-Efficient Split Federated Learning with Dynamic Feature Compression

Federated learning (FL) allows distributed clients in edge computing (EC) systems to collectively train a global model without sharing their local data. To ease the strain of training large-scale models on clients with limited resources, split federated learning (SFL) has been proposed as a viable alternative, combining both data and model parallelism. Nonetheless, the iterative transmission of intermediate features and gradients between clients and the server incurs considerable communication overhead. Additionally, the existing SFL methods struggle to effectively address the inherent system heterogeneity of edge clients, potentially leading to significant resource costs. To address these challenges, an efficient SFL method is proposed in this paper, termed FedDFC, which integrates dynamic feature compression. Specifically, FedDFC utilizes time proxy metrics to assess the heterogeneous capabilities of different clients and dynamically determine varying compression ratios for them. Furthermore, to prevent feature compression from affecting the model performance, a correction method is applied to the gradients related to the compressed features. Extensive experimental results demonstrate that FedDFC reduces the network traffic consumption by 14.7 % to 53.7 % while achieving comparable test accuracy to the baselines.