Decentralized Federated Learning in LEO-Satellite-Based IoE Communications: Latency Optimization Under Reliability Constraints

Artificial intelligence-empowered low-Earth orbit (LEO) satellite networks have the great potential to provide robust and ubiquitous communications capabilities for Internet of Everything (IoE) services. Due to the intrinsic high decentralization, it is expected for LEO satellite nodes to deploy decentralized federated learning (DFL) for model training collaboratively while preserving data privacy. However, the challenges lie in the latency management of the DFL process under communications reliability constraints. To address the issue, we propose a novel DFL framework that incorporates halving and doubling to balance the load of networks. Considering the dependency of aggregation, the latency of DFL clients is recursively derived. Under constraints of energy consumption and network reliability, an optimization problem is formulated that aims to minimize the mean latency, solved using the particle swarm optimization algorithm. Additionally, we introduce a latency-aware scheduling strategy to further improve DFL latency efficiency by leveraging overlapping intersatellite links among DFL clients. Simulation results show that the proposed methods significantly accelerate the DFL process by up to 16.2% and enhance its efficiency compared to the alternatives.