DSFL: Decentralized Satellite Federated Learning for Energy-Aware LEO Constellation Computing
Chenrui Wu, Yifei Zhu, Fangxin Wang
Abstract
The dense constellation of Low Earth Orbit (LEO) satellites plays a significant role in the sixth generation mobile network (6G) and Terrestrial-Aerial-Space network. With the communication network composed of the LEO satellites, ground stations and terminals, global network services can be accessed anywhere and anytime. The global communication coverage and precious data resource embraces new opportunities to integrate the computing resources and data in LEO satellite constellations for intelligent learning tasks, like carbon estimation, transportation surveillance, forest fire detection, etc. Federated learning (FL) stands out as a promising paradigm towards this goal with a balance in privacy preservation and data utilization. Traditional FL employs a centralized deployment, i.e., regarding the ground station as the server and satellites as clients, which however faces two challenges: 1) The risk of single point failure from the centralized server. 2) Slow convergences resulting from satellites’ intermittent communication. To tackle the challenges, in this paper, we propose a decentralized satellite federated learning, considering the limited communication traffic in satellite communication, the privacy of data and the efficiency of machine learning. Satellites collaborate in a decentralized way to reach a consensus on model parameters, overcoming the effects of data heterogeneity between satellites. Considering the scarce power capacity, we further design an energy-aware communication strategy to prolong satellites life and avoid communication congestion. Experiments on real-world datasets demonstrate that our framework can speed training process by 5–10 times, and requires only 1/3 to 1/6 communication energy cost.