POAA: A Partial Occlusion Adversarial Attack Method Against DNN-Based SAR Target Recognition

Deep neural network (DNN)-based synthetic aperture radar (SAR) target recognition has been proven to be vulnerable to deliberately designed adversarial perturbations, providing valuable insights into protecting critical targets from survival threats. However, most existing adversarial attack methods against DNN-based SAR target recognition focus on the image domain. The direct addition of irregular global perturbations to SAR target images lacks physical relevance. To address this issue, a partial occlusion adversarial attack method against SAR target recognition networks is proposed in this article. First, an optimization model for adversarial perturbation generation is constructed, involving the strategy of partial occlusion and the design of an objective optimization function. Then, an efficient solution is provided for the constructed optimization model by employing the particle swarm optimization algorithm in target region. Finally, the optimal perturbation position is determined, enabling the misclassification of SAR target recognition networks. This method establishes a paradigm for generating adversarial perturbations based on partial occlusion. Extensive experiments conducted on the moving and stationary target acquisition and recognition and ATRNet-SAR dataset confirm the efficacy of the proposed method.