ABSTRACT Mobile ad hoc networks (MANETs) are inherently vulnerable to selfish forwarding, collusion, stealthy oscillations, and resource‐draining attacks. Due to their decentralized, dynamic, and adversarial nature, existing anomaly‐based and deep‐learning intrusion detection systems rely on static training distributions and fixed payoff structures. This limits their adaptability under dynamic network conditions. This paper presents a secure and reliable operation of MANETs under adversarial environments. The model integrates a reinforcement learning–driven (RL‐driven), payoff‐adaptive, three‐action game‐theoretic (GT) framework. This combination helps learning both the utility structure and the optimal defense strategies for MANET nodes. The model employs online payoff learning, Boltzmann‐guided Q‐learning, and feature‐aware utility estimation to dynamically update security, throughput, energy, and latency trade‐offs. Experiments on a large‐scale NS‐3 mobility dataset ( trace snapshots) over 80‐node MANETs, 500 rounds, and 10 random seeds demonstrate a strong stability and resilience. The system achieves near‐optimal reliability (packet delivery ratio, , low latency , efficient energy usage controlled energy consumption and high average utility . Under seven adversarial threats, the framework consistently restores optimal operation. misreporting impact is reduced, with utility improving from , collusion inflation suppression remains , and regret normalization goes from to . Overall, the proposed GT–RL model provides adaptive equilibrium formation, multi‐metric flexibility, and theoretical convergence guarantees for secure MANET operation.
Zelani et al. (2026) studied this question.