SAR-GPA: SAR Generation Perturbation Algorithm

Abstract

The deep learning is widely used in optical image and synthetic aperture radar (SAR) image. Current academic research shows that adversarial perturbation can effectively attack the deep learning network in optical image. However, in SAR image target recognition network, the existence of universal perturbations and generation approach needs to be further explored. Here, this article firstly proposes a systematic SAR generation perturbation algorithm (SAR-GPA) for target recognition network. The modulation phase sequences of the jamming points can vary casually by using the state-of-the-art electromagnetic metasurface technology. Therefore, when it acts on the SAR deceptive jamming system, it can produce artificial controllable perturbations. First, we take the imperceptible perturbations from universal adversarial perturbations (UAP) as reference to construct a unconstrained minimum optimization problem to find the specific sequences. Then, we solve this issue by adaptive moment estimation (Adam) optimizer.Thus, the SAR adversarial examples can be quickly and flexibly generated through our system. Finally, We design a series of simulation and experiment to verify the effectiveness of the adversarial examples and also the modulation sequences. According to the results, different from the traditional SAR blanket jamming methods, our approach can quickly generate imperceptible jamming, which can effectively attack three classical recognition models.