对抗制
合成孔径雷达
深度学习
计算机科学
人工智能
欺骗
自动目标识别
图像(数学)
计算机视觉
心理学
社会心理学
作者
Yi Yu,Haiyan Zou,Fan Zhang
标识
DOI:10.1109/igarss52108.2023.10282390
摘要
The use of deep learning technology in Synthetic Aperture Radar Automatic Target Recognition (SAR ATR) has attracted significant attention. However, it has been noted that deep learning is vulnerable to adversarial attacks, and of course SAR ATR is no exception. Notably, several studies in the past two years have generated SAR adversarial examples, which can successfully deceive a specific well-trained SAR ATR deep network. Considering that research on optical adversarial attacks has progressed toward adversarial examples in the physical world, this paper aims to examine the feasibility of SAR adversarial examples in the physical world. Inspired by Google Sticker, we propose a SAR ATR adversarial-patch-based deception method, namely SAR sticker. The images crafted by our SAR sticker exhibit marked resemblance to their corresponding original images, yet they possess great potential in launching potent attacks on state-of-the-art SAR ATR models, achieving a 76% fooling rate.
科研通智能强力驱动
Strongly Powered by AbleSci AI