Satellite Federated Edge Learning: Architecture Design and Convergence Analysis

Satellite Federated Edge Learning: Architecture Design and Convergence Analysis

2 Apr 2024 | Yuanming Shi, Senior Member, IEEE, Li Zeng, Graduate Student Member, IEEE, Jingyang Zhu, Graduate Student Member, IEEE, Yong Zhou, Senior Member, IEEE, Chunxiao Jiang, Senior Member, IEEE, and Khaled B. Letaief, Fellow, IEEE
The paper introduces a novel Federated Edge Learning (FEEL) algorithm, named FEDMEGA, designed for low-earth-orbit (LEO) mega-constellation networks. FEDMEGA addresses the challenges of high mobility and short ground-to-satellite link (GSL) duration in LEO networks by integrating inter-satellite links (ISLs) for intra-orbit model aggregation. The algorithm reduces the usage of GSLs by increasing the frequency of intra-orbit training rounds and using a ring-all-reduced based intra-orbit aggregation mechanism. It also employs a network flow-based transmission scheme for global model aggregation, enhancing transmission efficiency. Theoretical convergence analysis is provided, and extensive simulations show that FEDMEGA outperforms existing satellite FEEL algorithms, achieving a 30% improvement in convergence rate. The algorithm's performance is evaluated on synthetic and real datasets, demonstrating better prediction accuracy and faster convergence.The paper introduces a novel Federated Edge Learning (FEEL) algorithm, named FEDMEGA, designed for low-earth-orbit (LEO) mega-constellation networks. FEDMEGA addresses the challenges of high mobility and short ground-to-satellite link (GSL) duration in LEO networks by integrating inter-satellite links (ISLs) for intra-orbit model aggregation. The algorithm reduces the usage of GSLs by increasing the frequency of intra-orbit training rounds and using a ring-all-reduced based intra-orbit aggregation mechanism. It also employs a network flow-based transmission scheme for global model aggregation, enhancing transmission efficiency. Theoretical convergence analysis is provided, and extensive simulations show that FEDMEGA outperforms existing satellite FEEL algorithms, achieving a 30% improvement in convergence rate. The algorithm's performance is evaluated on synthetic and real datasets, demonstrating better prediction accuracy and faster convergence.
Reach us at info@study.space