Adaptive Routing Strategy for Satellite Internet Under Smart Jamming Attacks via Graph Attention Network

Satellite internet has advanced rapidly, enabling global connectivity in remote regions and disaster relief. However, low Earth orbit (LEO) constellations, with dynamic topologies and periodic visibility, are highly vulnerable to smart jamming, which degrades reliability. To address this challenge, we propose an adaptive routing strategy that combines Graph Attention Networks (GAT) with Deep Reinforcement Learning (DRL). The constellation is modeled as a graph, and the routing task is formulated as a Markov Decision Process (MDP). GAT extracts topological features with multi-head attention to capture diverse inter-satellite dependencies, while the Actor-Critic DRL module optimizes next-hop selection using a reward function balancing delay, distance, and reliability. The framework adopts offline training and online decision-making, enabling rapid adaptation under jamming. Simulations on a Starlink constellation show that, compared with the exisiting baselines, the proposed GAT-DRL achieves lower routing delay and stronger robustness across different jamming intensities. Results demonstrate that integrating GAT's feature modeling with DRL's policy optimization significantly enhances anti-jamming capability and routing stability. This work provides a foundation for extending intelligent routing to multi-layer constellations and more complex interference scenarios, supporting resilient satellite internet infrastructure.