LEO Satellites Accelerating Edge Intelligence: Model Transfer and Similarity-Aware Initialization for Federated Learning

The vision of 6G networks is to realize a fully connected and intelligent world by integrating large-scale distributed intelligence with Federated Learning (FL) and the Internet of Things (IoT). However, conventional FL faces significant challenges in resource-constrained IoT environments, including slow convergence and high communication overhead during training operations. These challenges, combined with the stringent service requirements of IoT subsystems, highlight the need for more efficient FL mechanisms. Low Earth Orbit (LEO) satellites, with their global coverage and onboard edge computing capabilities, provide a promising platform for supporting large-scale distributed intelligence solutions. In this work, we propose a multi-layered Non-Terrestrial Network (NTN)-based FL framework that enables intelligent IoT applications. The framework leverages LEO satellite-assisted model search, transfer, and similarity-aware FL initialization to accelerate learning and reduce communication costs. We further analyze the trade-off between the overhead of model search and transfer over the LEO satellite network and the resulting improvements in FL convergence performance. Performance evaluations reveal that even a limited satellite-assisted search can cut convergence latency by nearly half compared to conventional FL, demonstrating the importance of LEO satellite-assisted model search, transfer, and similarity-aware FL initialization. These results highlight the potential of satellite-enabled FL as a key enabler of scalable, intelligent, and globally connected 6G IoT ecosystems.