Split Federated Learning-Empowered Energy-Efficient Mobile Traffic Prediction Over UAVs

In this letter, to alleviate the training burden over Unmanned Aerial Vehicles (UAVs) for generating the mobile traffic prediction model collaboratively, we design a novel energy-efficient mobile traffic prediction framework empowered by Split Federated Learning (SFL) for UAV networks, termed E-SFL. For this purpose, we rigorously formulated an analytical model of the overall energy consumption of UAVs, including both computing and networking energy consumption. The experimental results for two real-world mobile traffic datasets show that the proposed E-SFL surpasses previous state-of-the-art methods in terms of energy consumption with an acceptable accuracy loss.

I. Introduction

Cellular-connected unmanned aerial vehicle (UAV) networks are becoming integral components of sixth-generation (6G) networks, showcasing various applications (e.g., capture images and videos, and collect various types of sensing data) [1]. The rapid advancement of Artificial Intelligence (AI) technologies has led to AI-aided mobile traffic prediction schemes, previously explored for cost-efficient network resource management in terrestrial networks [2], [3], [4], [5], are now being adapted for UAV networks. By considering the distributed nature of mobile networks and privacy issues, one focus of recent studies is the use of Federated Learning (FL) [6] to create mobile traffic prediction models collaboratively. FL enables clients, such as multi-access edge computing (MEC) servers, to conduct local training and transmit local model parameters instead of entire datasets for server aggregation [7]. This approach helps reduce large transmission delays and processing bottlenecks associated with centralized learning (CL) servers. However, FL imposes a significant computational load on clients for local training [8]. This is particularly challenging when dealing with complex models like deep learning (DL), especially in UAV networks, where limited battery power and computational resources are major constraints.