IPAttack: imperceptible adversarial patch to attack object detectors

期刊论文
发表日期：

2025年2月19日

查看全文

Yongming Wen,
Peiyuan Si,
Wei Zhou,
Zongheng Zhao,
Chao Yi,
Renyang Liu

摘要

With the widespread application of deep learning, general object detectors have become increasingly popular in our daily lives. Extensive research, however, has shown that existing detectors are vulnerable to patch-based adversarial attacks, which fool such detectors by crafting adversarial patches. Although existing methods have made significant progress in terms of attack success rate, they still suffer from a highly perceptible problem, making it easy for humans to distinguish these evil examples. To address this issue, in this paper, we propose a novel spatial transform-based end-to-end patch attack method, called IPAttack, to synthesize imperceptible adversarial patches. Our approach estimates a flow field $\varvec{f}$ $f$ to formulate adversarial examples rather than introduce small $L_p$ $L_{p}$ -norm constrained external perturbations. Besides, to improve the imperceptibility and maintain a high attack performance, we propose the Object Detector Class Activation Map (OD-CAM) for objectors to extract the most interesting region, which will be applied to spatial transform to generate the final adversarial examples. Extensive experiments demonstrate that the proposed IPAttack can generate patch-wised adversarial examples with high imperceptibility while achieving the best attack performance compared to existing methods.