An Empirical Study Towards SAR Adversarial Examples

Adversarial attack and adversarial detection have become a hot issue in the field of deep learning based image forensics. However, current researches mainly focus on optical images. Synthetic aperture radar (SAR) images are quite different from the optical images in both imaging mechanism and data structure. This paper aims to study adversarial attack and adversarial detection for SAR images. Firstly, we analyze the distribution characteristics of SAR adversarial examples (AEs) in both output space and feature space by simply transferring optical attacks. In order to match the digital perturbation with the scattering energy of target, we then propose a generation method of SAR AEs with regional constraint. Experiments show that the proposed method decreases the attack performance of four classic adversarial attacks but leads to difficult detection. Finally, we point out an open issue that decreasing the perturbation scale leads to the degradation of adversarial detection against both optical AEs and SAR AEs.