Mobile ad hoc networks (MANETs) operating in hostile environments are vulnerable to coordinated wireless, routing, trust, and topology attacks. This paper proposes a tri-level
adversarial GNN–DRL framework that integrates adaptive graph attack generation, trust-aware topology recovery, hierarchical local-global graph encoding, and self-play reinforcement learning for joint route, relay, channel, and transmit-power control. The attacker models jamming, edge deletion, fake-edge and fake-node insertion, trust manipulation, and routing poisoning under a constrained attack budget. The defender reconstructs reliable graph structures before performing hierarchical communication control. Evaluation over five independent seeds shows that the proposed framework reaches a defender reward of 752.45 and reduces the attack success rate to 0.143. Under a 30% attack budget, it maintains a packet delivery ratio of 0.675, recovery accuracy of 0.751, and communication robustness index of 0.662, while limiting average delay to 155.36 ms. At 20 m/s, normalized connectivity and route stability remain 0.604 and 0.595, respectively. The results demonstrate improved
topology recovery, attack resilience, and communication stability in adversarial MANET conditions