Joint Client Selection and Model Compression for Efficient FL in UAV-Assisted Wireless Networks

Deploying federated learning (FL) applications in unmanned aerial vehicle (UAV)-assisted wireless networks can enable ground terminals (GTs) to perform complex machine learning tasks with their own data. However, the FL is inefficient in practice due to the massive model parameters and device heterogeneity. In this paper, we propose a joint client selection and model compression scheme for FL (csmcFL) to improve training efficiency. Specifically, the average throughput of users is first improved by optimizing the UAV deployment location based on user communication fairness. Then, a low-rank decomposition of the fully connected layer in the CNN is performed to compress the model parameters, and partial devices are screened to implement model compression through the client selection strategy to alleviate the excessive aggregation time due to device heterogeneity. We perform extensive simulation experiments in different data distribution scenarios, and the experimental results show that the proposed scheme significantly reduces the data volume of the transmitted model while achieving higher model accuracy compared to the baseline scheme.

I. Introduction

As One of the promising potential applications in the future of the sixth generation(6G) communication networks, unmanned aerial vehicle (UAV) can provide communication and computing services for ground users in disaster areas or temporary hot spots quickly by carrying base stations or servers under its inherent high mobility, high flexibility, and rapid deployment characteristics [1], while it is often impractical to deploy ground base stations in these areas. Correspondingly, UAV-assisted wireless communication has found wide use in data collection and machine learning model training [2], [3].