SAR Image Adversarial Samples Generation Based on Parametric Model
Xunwang Dang, Hua Yan, Liping Hu, Xuejian Feng, Chaoying Huo, Hongcheng Yin · 2021
At present, synthetic aperture radar (SAR) image recognition algorithm has achieved high accuracy. With the development of target recognition algorithm, the adversarial sample generation algorithms which can reduce the recognition performance have been gradually developed. To reduce the recognition performance of the sample, the original samples are adjusted slightly using certain algorithms. In this paper, a method of SAR image adversarial sample generation based on parametric model is proposed to ensure the physical feasibility of adversarial samples. First, the parametric model of camouflage structure is added to the original target physical model to realize the SAR image generation based on the parametric model. Then, an improved fast gradient sign method is used to generate SAR images of targets with camouflage structures. Numerical examples validate such method by reducing the accuracy of the original recognition algorithm.