返回
Low frequency sparse adversarial attack
DOI:10.1016/j.cose.2023.103379.png)
摘要
En 中文
Deep neural network (DNN) is found to be vulnerable to adversarial attacks. Gradient-based adversarial attacks usually craft perturbation on the whole pixels of input image, making them easier to be perceived or detected by the detection-based defenses, in the case of using large perturbation budget. In this paper, we present a new sparse approach to perturb image, which we call Robust Gradient-based Low Frequency search (RGLF). We show that RGLF provides a convenient way to modify any adversarial attack so that its imperceptibility can be improved significantly, not only to human eyes, but also to some popular detection-based defenses. Technically, RGLF first uses a robust gradient mask (RGM) to localize the semantic regions of input image, and then, applies a technique, called adaptive frequency search (AFS), to generate the low-frequency perturbations we expect, and finally, embeds the perturbations in the semantic regions. Extensive experiments are conducted to show that, i) the proposed RGLF technique can remarkably reduce the detection rates of the gradient-based adversarial attacks by 25%, 36%, and 30%, respectively for three SOTA detection-based defenses, namely, SBD, LiBRe, and SimCat; ii) The imperceptibility of the adversarial attacks modified by RGLF is improved significantly, even in the case of using large perturbation budget.
Keyword:
Adversarial attack
Imperceptible attack
Sparse attack
Low-frequency attack
Image classification
期刊
C
IF:
5.4
论文数:
4.6K
被引数:
1.8W
机构
引用论文
A Kinetic Model with Ordered Cytoplasmic Dissociation for SUC1, an Arabidopsis H + /Sucrose Cotransporter Expressed in Xenopus Oocytes在非洲爪蟾卵母细胞中表达的拟南芥H/蔗糖协同转运蛋白SUC1的有序细胞质解离动力学模型
Misleading attention and classification: An adversarial attack to fool object detection models in the real world误导性关注和分类: 在现实世界中欺骗对象检测模型的对抗性攻击
COMPUTERS & SECURITY
IF5.4
FE-DaST: Fast and effective data-free substitute training for black-box adversarial attacksFe-dast: 针对黑盒对抗攻击的快速有效的无数据替代训练
COMPUTERS & SECURITY
IF5.4
Detecting Adversarial Image Examples in Deep Neural Networks with Adaptive Noise Reduction基于自适应降噪的深度神经网络对抗图像样本检测
Gradient-based learning applied to document recognition基于梯度的学习在文档识别中的应用
PROCEEDINGS OF THE IEEE
IF25.9

