CSAFL: Communication-Efficient and Staleness-Aware Asynchronous Federated Learning for LEO Satellite Networks
The Satellite Internet of Things (SIoT) leverages low Earth orbit (LEO) satellite constellations to support wide-area, real-time, and intelligent services such as disaster monitoring, ship tracking, and environmental sensing. These tasks demand collaborative model training across multiple satellites to improve prediction accuracy and response efficiency. To support such multi-satellite collaboration, federated learning (FL) provides a natural solution by enabling satellites to train models locally and aggregate knowledge without requiring direct data sharing. However, applying FL in LEO satellite networks is challenging due to their high mobility and intermittent connectivity. The limited and short-lived uplink bandwidth restricts frequent model uploads. Asynchronous satellite availability disrupts the timing of global aggregation. Transmission delays also lead to stale model updates, which can degrade convergence and model stability. To address these issues, we propose Communication-Efficient and Staleness-Aware Asynchronous Federated Learning (CSAFL), a framework tailored for LEO satellite networks. CSAFL consists of three key components: 1) A gradient sparsification method uploads only important parameters to reduce transmission cost and adapt to limited bandwidth; 2) Each satellite performs asynchronous updates when a connection is available, which avoids delays caused by global synchronization; 3) The server adopts a staleness-aware weighted aggregation rule that considers the freshness of each update to improve convergence and model stability. We evaluate CSAFL on two benchmark datasets, and experimental results show that it consistently outperforms baseline methods, achieving over 9% higher accuracy and reducing communication cost by over 80% compared to FedAvg, FedAsync, and SparseFed. These findings demonstrate the effectiveness of CSAFL in highly dynamic and bandwidth-constrained LEO satellite networks.