5GC Enabled Lightweight Federated Learning under Communication Quality Constraint
Wenqi Zhang, Chen Sun, Xiaoxue Wang, Lantao Li, Tao Cui, Haojin Li · 2024
The ingenious training process of Federated Learning (FL) empowers multi-party collaborative training while avoiding the information privacy and security issues. However, there are differences in communication quality among FL clients. Particularly, for the FL clients with abysmal communication quality, the integral local models cannot be well uploaded to FL server. In this article, leveraging the Network Data Analytics Function (NWDAF) of 5 G Core Network (5GC), NWDAF Enabled Lightweight Federated Learning (NE-LFL) mechanism is proposed to solve the FL training under communication quality constraint. According to the model size and analysed communication quality, NWDAF analyses which FL clients cannot upload the integral model. Furthermore, to realize the customized lightweight model transmission for all FL clients, the significant model parameters will be extracted based on FL clients’ communication quality. The simulation results demonstrate the effectiveness and feasibility of the proposed NE-LFL mechanism. Specially, the proposed mechanism improves the model accuracy at most by $12.09 \%$ and $\mathbf{1. 5 5 \%}$ under Multi-Layer Perceptron (MLP) and Convolutional Neural Network (CNN), respectively.