SPAttack: A Physically Feasible Adversarial Patch Attack Against SAR Target Detection

Due to the inherent vulnerability of convolutional neural network (CNN), target detection networks are susceptible to adversarial perturbations. Although various attack algorithms can substantially impair the performance of detection networks, their effectiveness does not readily transfer to synthetic aperture radar (SAR) images due to the distinct imaging mechanism. A new strip-shaped adversarial patch attack named SPAttack is designed in this letter. The algorithm incorporates SAR attributed scattering centers for image reconstruction and filters scattering points to determine strong scattering regions as the perturbation area. The echo signal is weakened by introducing perturbations into strong scattering regions, leading to the disruption of SAR imaging. The SPAttack patch is restricted to SAR target regions and scaled to match the target size rather than pixel-level perturbations, ensuring greater feasibility for real-world deployment. In the experiment, SPAttack is compared with several advanced attack algorithms. The results show that SPAttack achieves excellent attack performance while being physically realizable and computationally efficient. Compared with other existing methods, SPAttack improves the attack success rate by 3%.