GENERATING ADVERSARIAL EXAMPLES ON SAR IMAGES BY OPTIMIZING FLOW FIELD DIRECTLY IN FREQUENCY DOMAIN

Deep neural networks (DNNs) have become significant methods for SAR image analysis. However, there is a non-negligible security problem with deep learning. DNNs are particularly vulnerable to adversarial perturbations added to the input images. The study of adversarial examples generation can undoubtedly facilitate the study of defense algorithms and effectively improve the security and robustness of deep learning systems. We propose a novel method for generating adversarial examples on SAR images by optimizing flow field directly in frequency domain. We first transform the SAR images from spatial domain to frequency domain via discrete cosine transform. Then we optimize the flow field directly in frequency domain to generate the adversarial examples. Experiment of classification task on MSTAR dataset shows that our method can fool the DNN with 100% success rate, and the average number of attacks is greatly reduced from 36.58 to 5.58, which significantly improves the attack efficiency.