MIGAA: A Physical Adversarial Attack Method against SAR Recognition Models
Jianyue Xie, Bo Peng, Zhengzhi Lu, Jie Zhou, Bowen Peng · 2024
Deep neural networks (DNNs) utilized in SAR-TAR systems have been shown their vulnerability to adversarial attacks. However, the current SAR attack methods rely heavily on external information and overlook the physical realizability of adversarial examples, which raises concerns about their practical feasibility. To tackle this problem, we introduce a novel physically feasible black-box attack, named Metasurface Interfer-ence Guided Adversarial Attack (MIGAA). Our approach aims to generate adversarial samples by leveraging time-modulated metasurface technology, resulting in practical perturbations applicable in real-world scenarios. Specifically, we employ Particle Swarm Optimization (PSO) and multiple enhanced strategies to effectively craft metasurface-based adversarial perturbations. Additionally, due to the efficiency of the PSO algorithm, our method successfully misleads state-of-the-art DNN models in black-box settings with a limited number of queries. Experimen-tal results on the MSTAR dataset demonstrate that the proposed MIGAA outperforms existing black-box attacks using substitute models and achieves the highest physical feasibility in SAR-ATR systems.