STELLAR: Similarity-Based Satellite Federated Learning for Malicious Traffic Recognition

Resource constraints and data heterogeneity pose significant hurdles for malicious traffic detection in satellite networks. To address this, we propose STELLAR, a similarity-based federated learning framework tailored for efficient space computing. STELLAR introduces a multi-dimensional similarity metric to dynamically select representative nodes, effectively eliminating computational redundancy. Furthermore, it ensures system robustness and trust through a time-window asynchronous protocol with staleness compensation and a lightweight proxy-based authentication scheme. Evaluations demonstrate that STELLAR outperforms specialized baselines, reducing network-wide energy consumption by 77%–80% while achieving 99.72% detection accuracy in heterogeneous Non-IID environments. These results validate STELLAR as a sustainable and robust solution for distributed security in resource-constrained satellite networks.