Bait Tactics: Misleading DRL-Based Cognitive Anti-Jamming Communications via Adversarial Learning

Reinforcement learning (RL) has emerged as a promising tool for enabling adaptive and autonomous behavior in cognitive wireless networks, where online decision-making is required under partial or no prior knowledge of the environment. By interacting with the wireless environment, RL-based cognitive links can dynamically optimize transmission strategies to mitigate interference and improve throughput. However, this adaptability introduces a critical vulnerability as RL agents are inherently sensitive to deception. The present contribution exposes this vulnerability by analyzing a deep RL-based anti-jamming link that enhances its performance by exploiting jamming energy through intelligent use of backscattering and energy harvesting. We counter this advantage by proposing a deception-based jamming strategy, termed "bait tactics," which manipulates the agent’s perceived state transitions and increases reward variance through targeted environment manipulations. Simulation results demonstrate that the proposed adversarial strategy can reduce the victim’s throughput by up to 72% while conserving up to 67% of jamming power compared to a standard reactive jamming in the considered scenario. These findings reveal a critical vulnerability in RL-driven cognitive networks and highlight the need for robust, deception-resilient learning frameworks in future wireless systems.