STARFed: Link-Aware Defense Against Poisoning Attacks in Satellite-Terrestrial Federated Learning

Satellite-ground integrated computation where ma chine learning models trained on satellites and aggregated on Earth offers novel opportunities for federated learning (FL). While satellites in space provide isolated computing environments, satellite-terrestrial (S-T) communication links are exposed to spoofing and hijacking attacks, making transmitted models vulnerable to poisoning attacks. To address this paradigm specific threat, we introduce STARFed, a novel framework that enhances robustness of satellite-based FL by leveraging S-T link characteristics during model transmission. It comprises three components: (1) crowdsourcing-based link authentication, (2) hybrid poison model detection based on both S-T link and model characteristics, and (3) reputation-based model filtering against adaptive adversaries. Our link-aware defense is of independent interest and can be combined with various FL robust aggregation schemes. We evaluate the framework's resilience through com prehensive experiments spanning five dataset-model settings and five attacks, including both model and data poisoning attacks. The framework's performance is compared with six state-of-the-art robust FL aggregation schemes in scenarios with varying degrees of non-IID data distribution, client dropout, and adversarial participation. STARFed demonstrates robust performance across all test scenarios, standing as the only defense mechanism to maintain effectiveness throughout. In the most favorable case, it achieves an increase in FL accuracy of 15.6% compared to the best link-unaware aggregation scheme, with minimal overhead introduced.