An Adversarial Example Generation Method Based on Mask Extraction

The traditional perturbation generation technique in the process of adversarial example generation requires a large amount of perturbation, which makes the example easy to be detected. An efficient adversarial example generation method for Synthetic Aperture Radar (SAR) images based on mask extraction is proposed to this issue. This method utilizes mean filtering, threshold segmentation, and dilation to design a mask module that extracts the main target regions in SAR images. Different subject perturbation coefficients and background perturbation coefficients are set. And the strength of perturbations in the main and background parts of adversarial samples are adjusted by Non-dominated Sorting Genetic Algorithm II (NSGA-II). It produces for a trade-off between the attack success rate (ASR) and the structural similarity index measure (SSIM) in the generation of adversarial samples, ensuring the overall effectiveness of the system. Simulation experiments show that compared to traditional attack methods, the efficient adversarial example generation method for SAR images based on mask extraction can maintain a high structural similarity index measure while effectively increasing the attack success rate without degrading image quality. In addition, through the adaptive perturbation parameter setting of NSGA-II, the optimal solution of both the attack rate and SSIM can be ensured.