Hu Jincheng, Li Jihao, Hou Zhuoran, Jiang Jingjing, Liu Cunjia, Chu Liang, Huang Yanjun, Zhang Yuanjian
Department of Aeronautical and Automotive Engineering, Loughborough University, Leicestershire LE11 3TU UK.
College of Automotive Engineering, Jilin University, Changchun 130022 China.
iScience. 2023 Jul 13;26(9):107393. doi: 10.1016/j.isci.2023.107393. eCollection 2023 Sep 15.
Severe weather conditions pose a significant challenge for computer vision algorithms in autonomous driving applications, particularly regarding robustness. Image rain-removal algorithms have emerged as a potential solution by leveraging the power of neural networks to restore rain-free backgrounds in images. However, existing research overlooks the vulnerability concerns in neural networks, which exposes a potential threat to the intelligent perception of autonomous vehicles in rainy conditions. This paper proposes a universal rain-removal attack (URA) that exploits the vulnerability of image rain-removal algorithms. By generating a non-additive spatial perturbation, URA significantly diminishes scene restoration similarity and image quality. The imperceptible and generic perturbation employed by URA makes it a crucial tool for vulnerability detection in image rain-removal algorithms and a potential real-world AI attack method. Experimental results demonstrate that URA can reduce scene repair capability by 39.5% and image generation quality by 26.4%, effectively targeting state-of-the-art rain-removal algorithms.
恶劣天气条件对自动驾驶应用中的计算机视觉算法构成了重大挑战，尤其是在鲁棒性方面。图像去雨算法通过利用神经网络的能力来恢复图像中无雨的背景，已成为一种潜在的解决方案。然而，现有研究忽略了神经网络中的脆弱性问题，这对雨天条件下自动驾驶车辆的智能感知构成了潜在威胁。本文提出了一种通用去雨攻击（URA），该攻击利用了图像去雨算法的脆弱性。通过生成非加性空间扰动，URA显著降低了场景恢复相似度和图像质量。URA所采用的难以察觉且通用的扰动使其成为图像去雨算法中脆弱性检测的关键工具以及一种潜在的现实世界人工智能攻击方法。实验结果表明，URA可以将场景修复能力降低39.5%，将图像生成质量降低26.4%，有效地针对了最先进的去雨算法。
