Impact of Network Dynamics on Decentralized Federated Learning in LEO Satellite Computing
Sainan Shi, T. Liu, Ruyun Zhang · 2024
The rapid development of Low Earth Orbit (LEO) satellites is transforming satellite computing, creating new opportunities and challenges for global communication and data processing. While LEO satellites offer seamless communication coverage and low latency, their high velocity and limited transmission bandwidth make centralized data processing impractical. Federated Learning (FL) distributes learning across nodes but relies on central ground servers, posing risks of single point failures and extensive communication burdens. Decentralized Federated Learning (DFL) emerges as a promising alternative, eliminating the need for central servers by enabling direct model exchanges between neighboring satellites. This approach enhances robustness and scalability by reducing single-point failures and avoiding the need for a known global network topology. However, it should be noted that the performance of DFL is greatly affected by the network topology. This study offers the first investigation into the impact of dynamic network topologies on DFL from the perspective of graph centrality. Specifically, it is observed that an increased satellites communication range can enhance connectivity between nodes and improve the performance of DFL. However, no noticeable relationship is observed between the LEO speed and the connectivity between nodes, nor in the performance of DFL.