Adversarial Examples for CNN-Based SAR Image Classification: An Experience Study

摘要

Synthetic aperture radar (SAR) has all-day and all-weather characteristics and plays an extremely important role in the military field. The breakthroughs in deep learning methods represented by convolutional neural network (CNN) models have greatly improved the SAR image recognition accuracy. Classification models based on CNNs can perform high-precision classification, but there are security problems against adversarial examples (AEs). However, the research on AEs is mostly limited to natural images, and remote sensing images (SAR, multispectral, etc.) have not been extensively studied. To explore the basic characteristics of AEs of SAR images (ASIs), we use two classic white-box attack methods to generate ASIs from two SAR image classification datasets and then evaluate the vulnerability of six commonly used CNNs. The results show that ASIs are quite effective in fooling CNNs trained on SAR images, as indicated by the obtained high attack success rate. Due to the structural differences among CNNs, different CNNs present different vulnerabilities in the face of ASIs. We found that ASIs generated by nontarget attack algorithms feature attack selectivity, which is related to the feature space distribution of the original SAR images and the decision boundary of the classification model. We propose the sample-boundary-based AE selectivity distance to successfully explain the attack selectivity of ASIs. We also analyze the effects of image parameters, such as image size and number of channels, on the attack success rate of ASIs through parameter sensitivity. The experimental results of this study provide data support and an effective reference for attacks on and the defense capabilities of various CNNs with regard to AEs in SAR image classification models.