Reinforcement Learning-Based Adversarial Scatterer Generation for Attacking SAR Image Target Detectors

Adversarial attacks have been extensively studied in the context of synthetic aperture radar (SAR) target recognition. Recent studies has further revealed that SAR target detection models are also vulnerable to adversarial examples. However, most existing attack methods targeting detection models rely on gradient-based optimization strategies, which introduce global pixel-level perturbations without considering physical realizability. To address this issue, this work proposes a reinforcement learning-based adversarial scatterer generation method (RL-ASGM) that incorporates SAR imaging domain knowledge. The core idea is to deceive detectors by strategically superimposing carefully designed adversarial scatterers within target regions. The RL-ASGM framework consists of two key components: a reinforcement learning-based adversarial scatterer parameter optimizer (RL-ASPO) and an agent-cascaded black-box querying framework (ACQF). For RL-ASPO, the attributable scattering center model (ASCM) is adopted to characterize the scattering behavior of typical geometric structures, and reinforcement learning algorithms are utilized to train agents searching for effective scatterer parameters. To address multi-target scenario, ACQF employs a cascaded attack strategy that assigns a dedicated agent to each target. These agents iteratively query the detector and sequentially generate and deploy adversarial scatterers. Experimental results on the SAR-AIRCraft-1.0 and SIVED datasets demonstrate that the proposed method achieves an average attack success rate exceeding 80% across various mainstream detectors, while occluding no more than 65 pixels per image, validating its effectiveness and superiority. The code is available at https://github.com/bobp26/RL-ASGM.