A Target-region-based SAR ATR Adversarial Deception Method
Tianying Meng, Fan Zhang, Fei Ma · 2022 7th International Conference on Signal and Image Processing (ICSIP) · 2022
In order to attack the hostile deep learning based Synthetic Aperture Radar Automatic Target Recognition (SAR ATR) models, a lot of algorithms have been proposed to create adversarial examples by adding perturbations to random positions in the images. Practical applications prefer that these perturbations locate in the target regions, so that the adversarial examples can guide us to camouflage the targets by adding strong scattering objects or absorbing material on the targets. Regarding this weakness, a new target-region-based SAR ATR adversarial deception method is proposed in this paper. We first utilize the Gabor feature-based texture segmentation (GFTS) method to extract the mask of targets in SAR images. Then the mask parameters are introduced into the loss function of perturbation generator to aggregate the perturbations of SAR adversarial samples into the target regions. The experimental results show that our method can not only generate the adversarial samples effectively fooling the SAR target recognition networks but also the perturbation only appears in part of the target regions. It provides a theoretical basis for the disguise of SAR targets in the physical world.