A-MATD3: adversarial multi-agent twin delayed DDPG for resilient routing in SAGIN under intelligent jamming attacks

Space-Air-Ground Integrated Networks (SAGIN) have emerged as a pivotal architecture for enabling ubiquitous global connectivity in next-generation wireless systems. However, SAGIN impose routing requirements that differ fundamentally from those of relatively homogeneous and slowly varying terrestrial networks. Cross-layer routes traverse satellite, aerial, and terrestrial links with different mobility timescales, bandwidth, propagation distance, resource availability, and interference vulnerability. Consequently, the feasible next-hop set varies over time, routing nodes operate with partial local information, and traffic concentration can be exploited by an adaptive jammer. We therefore formulate end-to-end SAGIN routing under intelligent link jamming as a partially observable stochastic game between distributed cooperative forwarding agents and an adaptive antagonist. In this paper, routing specifically refers to end-to-end multi-hop service routing across the space, air, and ground layers, where heterogeneous tasks generated by ground service endpoints are forwarded through dynamically selected satellite, aerial, and terrestrial relay nodes. Traditional routing protocols fail to adapt to such complex environments, while existing deep reinforcement learning (DRL)-based approaches primarily optimize performance in benign conditions, largely neglecting robustness against strategic adversaries. To address this critical gap, we instantiate the robust SAGIN routing problem as a conservative two-player zero-sum stochastic game and propose a novel Adversarial Multi-Agent Twin Delayed Deep Deterministic Policy Gradient (A-MATD3) framework. The method adopts a centralized training with decentralized execution (CTDE) paradigm, integrating a hybrid Graph Convolutional Network (GCN)-Transformer encoder to effectively capture complex spatio-temporal dependencies in dynamic networks. But the contribution is not a new GCN, Transformer, or TD3 operator in isolation. Instead, we develop a routing-specific adversarial MARL framework that combines a time-varying cross-layer graph, spatio-temporal local observations, dynamically masked next-hop actions, heterogeneous task-aware rewards, and hybrid link-selection/power-allocation actions for the jammer. Extensive simulations show that the A-MATD3 achieves higher benign and attack-time performance than five reference routing algorithms trained according to their non-adversarial protocols. A controlled ablation further shows that removing antagonist-based training substantially reduces the robustness index (the robustness index decreases from 0.826 to 0.512). Under severe intelligent jamming, it achieves up to 48.4% higher task success rate and 35.2% lower end-to-end delay than the best-performing baseline, while maintaining superior throughput and fairness in attack-free conditions.