SAR-PeGA: A Generation Method of Adversarial Examples for SAR Image Target Recognition Network
Weijie Xia, Zhe Liu, Yi Li
Abstract
Deep learning (DL) is widely used in automatic target recognition (ATR) of synthetic aperture radar (SAR) images. Related researches show that DL models for SAR ATR are vulnerable to adversarial examples attack in the digital domain. However, how to generate adversarial examples in practical scenarios is critical and challenging. In this paper, we propose a systematic SAR perturbation generation algorithm (SAR-PeGA) for target recognition network. Firstly, assuming that some reflection phase tuning samples are located in the fixed area of SAR target, we adjust the phase characteristics of reflected signal with variable phase sequences. Secondly, we take the imperceptible perturbations from universal adversarial perturbations (UAP) as reference. Then, we construct the unconstrained minimum optimization model to find the specific phase sequences of tuning samples, and optimize the model with the adaptive moment estimation (Adam) optimizer. Finally, SAR adversarial examples can be flexibly generated through the proposed deceptive jamming model. Experimental results demonstrate that the proposed method can generate imperceptible jamming and effectively attack three classical recognition models.