Adversarial Attacks Against Deep Reinforcement Learning-Based Jamming Decision-Making

The intelligence of jamming decision-making has been significantly enhanced by deep reinforcement learning (DRL), owing to its capability to autonomously discover optimal strategies through interaction with the environment. This advancement creates an urgent demand for innovative antijamming techniques specifically designed to counter intelligent jammers. To this end, this work presents a black-box framework against intelligent jamming decision-making, in which adversarial attacks are employed to compromise the effectiveness of DRL-based jamming strategies. The core concept involves manipulating the opponent’s observations through deliberate adjustments of radar pulse parameters, thereby misleading its decision-making process. Specifically, we first introduce parameter-guided behavior cloning (PBC) to address the lack of prior knowledge about the victim agent. PBC trains a surrogate agent from demonstration data while aligning its parameters with those of the victim agent to approximate the target policy, thereby providing the gradient information required for updating adversarial perturbations. On this basis, we develop the boundary-constrained adaptive distortion attack (BADA), which incorporates an adaptive moment-based update mechanism to stabilize the search for optimal perturbations. An adaptive step size scheme is introduced, employing larger updates in the early phase to speed up convergence and smaller ones later to mitigate oscillations and stabilize the search around the optimum. To ensure physical feasibility, BADA strictly enforces radar pulse constraints by clipping parameters to their prescribed ranges at each iteration. Extensive experiments demonstrate that the proposed adversarial pulse parameters effectively disrupt DRL-based jammers, thereby improving radar detection performance under jamming conditions. These results validate the feasibility and effectiveness of the proposed antiintelligent jamming framework.