Federated Learning-Based Cross-layer Security Design for Satellite Networks

The extensive coverage of satellite networks robustly supports federated learning (FL) in multiple domains. This combination protects user privacy and enables extensive data training, with promising applications in remote healthcare, smart agriculture, and environmental monitoring. However, existing FL primarily focuses on data training and aggregation, with less attention given to the secure transmission of model data during upload and download processes. This paper explores cross-layer security in satellite networks, focusing on the physical and application layers. We propose a beamforming optimization scheme based on unsupervised neural network to guarantee secure transmissions without compromising FL training performance. Simulation results underscore the efficacy of our approach in securing physical layer transmissions and affirm its practicality in maintaining robust FL training outcomes.