Joint Data Allocation and LSTM-Based Server Selection With Parallelized Federated Learning in LEO Satellite IoT Networks

Abstract

Low earth orbit (LEO) satellite networks have emerged as a promising field for distributed Internet of Things (IoT) devices, particularly in latency-tolerant applications. Federated learning (FL) is implemented in LEO satellite IoT networks to preserve data privacy and facilitate machine learning (ML). However, the user who spends the longest time significantly hampers FL efficiency and degrades the Quality-of-Service (QoS), potentially leading to irreparable damage. To address this challenge, we propose a joint data allocation and server selection strategy based on long short-term memory (LSTM) with parallelized FL in LEO satellite IoT networks. Herein, data-parallel learning is utilized, allowing multiple users to collaboratively train ML networks to minimize latency. Moreover, server selection takes into account signal propagation delays as well as traffic loads forecasted by an LSTM network, thereby improving the efficiency even further. Specifically, the strategies are formulated as optimization problems and tackled using a line search sequential quadratic programming (SQP) method and a multiple-objective particle swarm optimization (MOPSO) algorithm. Simulation results show the effectiveness of the proposed strategy in reducing total latency and enhancing the efficiency of FL in LEO satellite IoT networks compared to the alternatives.