Secure inter-satellite AirComp with temporal channel prediction and unfolded graph neural network

Low Earth orbit (LEO) satellites, as the critical infrastructure for 6G space-air-ground integrated Networks, face the dual challenges of efficient massive data aggregation and security provisioning. Over-the-air computation (AirComp) leverages wireless waveform superposition to achieve the integration of computation and transmission. While it significantly enhances communication and computation efficiency, its inherent wireless broadcast nature makes it highly vulnerable to eavesdropping attacks. To address these issues, this paper proposes a deep learning-based secure AirComp framework that incorporates channel prediction and secure computation decision-making, targeting highly dynamic inter-satellite links under eavesdropping threats. First, based on prior knowledge of satellite orbital dynamics, a physics-driven channel evolution model is established. A temporal Transformer-based channel prediction network is designed, employing a residual learning paradigm to capture the nonlinear evolution of channel physical parameters from historical observations, thereby achieving high-precision channel prediction. On this basis, a deep unfolding graph neural network (GNN)-based AirComp policy solver is proposed. This solver maps iterative optimization procedures into multi-layer unfolded neural networks and models the interaction and collaboration among different satellite transmitters via message-passing mechanisms, efficiently generating eavesdropping-resilient secure AirComp beamforming strategies. Simulation results demonstrate that the proposed scheme significantly reduces prediction errors, compared to traditional methods. Furthermore, the proposed method achieves near-optimal secure AirComp rate under various network configurations, significantly outperforming existing heuristic baseline schemes.