Template-Based Universal Adversarial Perturbation for SAR Target Classification

Abstract

With deep learning models been widely applied to SAR images interpretation, the adversarial robustness problem of models has drawn much attention. Due to the application in military fields, SAR image processing have extremely requirements on security. However, the popular optimized adversarial perturbation algorithms calculated the perturbation for one single image at a time, thus cannot be adapted to real-time adversarial scenarios. Also, the perturbation is designed in full digital space and lacks of sematic information of the targets. Faced with those problems, this paper proposes a SAR universal adversarial perturbation (SAR-UAP) for SAR images adversarial attacks based on template-based lightweight generative model without discriminator. The model has simple structure which consists of six convolutional layers and uses the fixed template as the input of the generator. Template-based UAP ensures that once training is finished, the disturbances remain unchanged in spite of the changing of images chosen to attack. Model performance was tested on the MSTAR dataset for eight classical CNNs. Compared with other GAN-based UAP methods, SAR-UAP gained slightly higher fooling rate which is over 85%, while the illustration showed SAR-UAP are much more target-focused. The transferability of perturbation was also verified.