Robust Resource Allocation for V2V Communications Against Evolving Jamming

Abstract

Vehicle-to-vehiclecommunication establishes essential connectivity among intelligent connected vehicles, forming the decentralized wireless networks that underpin intelligent transportation systems (ITS). Such networks are inherently vulnerable to jamming attacks that disrupt critical message dissemination, significantly degrading system reliability and road safety. Most existing methods employ reinforcement learning (RL) to enhance wireless security, however, they often treat attackers as static components, overlooking the dynamic evolutionary capabilities of adversaries results in strategies that are prone to exploitation by evolving attackers. To bridge this research gap, we propose a curriculum-enhanced fictitious play (CEFP) RL method to counter dynamically evolving adversarial attacks. In such adversarial settings, the simultaneous training of defenders and attackers with conflicting optimization objectives causes their updates to mutually influence each other, which induces strategy cycling and prevents stable convergence to an equilibrium. To address this instability, CEFP introduces a population-based fictitious play mechanism that empowers the defender to cope with dynamic scenarios. While fictitious play stabilizes training, learned policies may still suffer from poor generalization against unseen attackers. To mitigate the consequent risk of catastrophic degradation of defense effectiveness, we introduce a reward-based curriculum learning mechanism that adaptively selects appropriate opponents, thereby facilitating the acquisition of robust policies. The experimental results show that our method maintains substantial total and per-vehicle packet reception rates in scenarios where attackers dynamically evolve, ensuring the reliability and security of ITS.