Outsmarting the Smart: Intelligent Jamming Strategies Against AI-Empowered Anti-Jamming Frameworks

Reinforcement learning (RL) has become a key enabler for realizing adaptive and autonomous decision-making in next-generation AI-driven wireless networks, enabling real-time optimization of transmission strategies to counter jamming attacks. However, this adaptability also introduces critical vulnerabilities since the reliance of RL agents on environmental feedback renders them susceptible to deception, particularly when adversaries manipulate the environment in order to mislead the learning process. Yet, even though prior research considered adversarial jamming with white or grey-box access to the RL agent, the challenge of black-box jamming, where the jammer adapts without explicit feedback on its impact, remains largely unexplored. With this motivation, the present contribution addresses a practical adversarial scenario where a smart anti-jamming agent does not just resist jamming but actively exploits jamming signals to increase its throughput, especially as jamming attacks intensify. Defeating such an adaptive agent is particularly challenging in black-box settings, where the jammer has no knowledge of the link’s internal mechanisms or reward structure. In this context, we systematically benchmark a variety of advanced reactive jamming strategies, including both interaction-driven and optimization-driven approaches, under these realistic constraints. The achieved results indicate that adaptive, learning-driven jammers can reliably force even intelligent anti-jamming links into suboptimal operation, causing substantial throughput loss while consuming significantly less jamming power than conventional reactive jamming attacks. These findings reveal a fundamental vulnerability in RL-driven cognitive networks and highlight the urgent call for more resilient learning frameworks in order to secure next generation wireless systems.