SCMA: A Scattering Center Model Attack on CNN-SAR Target Recognition
Weibo Qin, Bo Long, Feng Wang
Abstract
Convolutional neural networks (CNNs) have been widely used in SAR (Synthetic Aperture Radar) target recognition, which can extract feature automatically. However, due to its own structural flaws, CNNs are easy to be fooled by adversarial examples, even if they have excellent performance. In this letter, a novel attack named scattering center model attack (SCMA) is designed, and its generation process does not rely on the prior knowledge of any neural network. Therefore, we can get a stable way which can be applied to any neural network. In addition, an improved scattering center model extraction method, which is the pre-part of SCMA, can filter out the useless noise to optimize the stability of attack. In the experiment, SCMA is compared with advanced attack algorithms. From the experimental results, it is clear to find that SCMA has excellent performance in terms of transfer attack success rate. Furthermore, visualization and interpretability analysis underpin the theoretical feasibility of SCMA.