RAFL: Reputation-Aware Federated Learning with hierarchical aggregation in LEO satellite networks

Abstract

The dynamic observation data of the Earth from Low Earth Orbit (LEO) satellites can be used for weather analysis, urban planning, etc. The existing satellite federated learning (FL) schemes have two main drawbacks. First, most of schemes rely on frequent communication between satellites and ground station. Due to the transient and unstable communication links, significant convergence delays are observed. Second, the possibility of unreliable and malicious participants preventing model convergence and degrading performance is not considered. To address the above issues, a secure and efficient FL scheme RAFL for LEO constellations is proposed. RAFL employs aerial networking and hierarchical FL to reduce interactions between satellites and ground station, thereby enhances training efficiency and model performance. A self-adaptive reputation evaluation system is designed that incorporates malicious behavior detection, historical behavior analysis, and performance and contribution assessments of LEO satellites. This system can detect and resist the actions of unreliable and malicious participants, and select appropriate leaders, thereby enhancing the robustness of FL. Extensive experiments show that RAFL not only effectively resists poisoning attacks but also improves the convergence time of FL while maintaining the accuracy of the global model.