Joint Server Selection and Handover Design for Satellite-Based Federated Learning Using Mean-Field Evolutionary Approach

With the increasing development of Low-Earth-Orbit (LEO) Satelite Communications (SatComs), it is foreseen that they will play an important role in broadening the horizon of Federated Learning (FL). Specifically, SatComs can amplify FL by providing consistent global transmission, bridging terrestrial network gaps, and ensuring robust, reliable connectivity in remote or challenging terrains. In this paper, we consider a SatComs-based FL framework, where satellites in the low-earth orbit collaborate to serve as global servers, able to collect and aggregate FL model parameters transmitted from the mobile devices on the ground continuously. We investigate the joint server selection and handover design optimization problem for SatComs-based FL from the perspective of the system's energy consumption and performance. To address the scalability issue, we propose a Mean-Field-Evolutionary (MFEv) approach that simplifies the interaction between mobile devices as a distribution over their state space, known as the mean-field approximation. It iteratively updates devices' strategies along the Fokker-Planck gradient flow in satellites' strategy space. Our approach's complexity is linear in the number of mobile devices, and we prove that it converges to a unique optimal solution. Numerical simulations demonstrate that our approach is effective in economically benefiting the system and reducing algorithm running time.

I. Introduction

The continually increasing data generated at edge devices stems from billions of interconnected Internet-of-Things (IoT) devices as every active IoT client gathers its observed information and forwards it to the edge. Traditional Machine Learning (ML) methods typically aggregate the gathered data on a central data center or a single machine, and such a centralized learning scheme is common among AI-driven companies such as Google and Microsoft [1]. To enhance model performance using this centralized approach, users might have to sacrifice their privacy by transmitting personal data to these centers. This training approach can be intrusive to privacy, especially when individuals have to share personal or sensitive information to improve the training model's performance. However, the emergence of Federated Learning (FL) happened to overcome these challenges. FL is an innovative approach to centralized ML where a model is trained across multiple IoT devices or servers while keeping the data localized [2]. Instead of sending the data to a central server for training, FL sends the model to the IoT devices. Each device computes an update to the model based on its local data, and then sends only this update back to the central server, where it is aggregated with updates from other devices to improve the global model.