NTN-Aware Federated Learning Framework With Hierarchical Satellite Aggregation and LoRA-Based Parameter Efficiency
Nonterrestrial networks (NTNs) based on low-Earth-orbit (LEO) satellite constellations provide promising platforms for conducting global-scale federated learning (FL). However, a fundamental feasibility barrier remains: transmitting full-model updates often exceeds the typical LEO visibility windows (30–90 s), resulting in systematic client dropouts and unstable training processes. The existing approaches largely treat communication efficiency and satellite topologies independently, leaving this challenge unresolved. In this paper, we propose a unified NTN-aware FL framework that integrates low-rank adaptation (LoRA) with a three-tier hierarchical aggregation architecture. We show that LoRA resolves the feasibility barrier not only through incremental compression but also by reducing the uplink payload size by <inline-formula> <tex-math notation="LaTeX">$74.6\times $ </tex-math></inline-formula>, thereby shifting the system bottleneck from communication-limited operations to computation-limited operations. This shift enables a hierarchical aggregation scheme that is otherwise infeasible under LEO visibility constraints. The proposed three-tier protocol combines satellite-side weighted aggregation, intersatellite link (ISL)-averaging consensus, and gateway-level global aggregation. To address intermittent connectivity issues, we further develop a staleness-aware asynchronous extension with a satellite-tailored discount function. In addition, personalized LoRA adapters enable client-specific adaptations to be implemented under heterogeneous channel conditions. We establish a rigorous NTN system model that captures topology dynamics and visibility constraints and prove its convergence under standard nonconvex assumptions. Simulations performed under realistic NTN settings demonstrate that the proposed method achieves <inline-formula> <tex-math notation="LaTeX">${R} ^{2}$ </tex-math></inline-formula> = 91.3%, closely matching full-model FL (91.8%) with a <inline-formula> <tex-math notation="LaTeX">$74.6\times $ </tex-math></inline-formula> parameter reduction. In terms of latency, the uplink contribution is reduced from 55.8% to 4.0%, while client-side computations account for 93.8% of the end-to-end latency. In terms of reliability, the client dropout rate is lowered from 65.3% to 25.4%. Regarding efficiency, ISL traffic is reduced by <inline-formula> <tex-math notation="LaTeX">$23\times $ </tex-math></inline-formula> while achieving sublinear round completion scaling up to 10,000 clients.