Patch is enough: naturalistic adversarial patch against vision-language pre-training models

Abstract Visual language pre-training (VLP) models have demonstrated significant success in various domains, but they remain vulnerable to adversarial attacks. Addressing these adversarial vulnerabilities is crucial for enhancing security in multi-modal learning. Traditionally, adversarial methods t...

Full description

Saved in:
Bibliographic Details
Main Authors: Dehong Kong, Siyuan Liang, Xiaopeng Zhu, Yuansheng Zhong, Wenqi Ren
Format: Article
Language:English
Published: Springer 2024-12-01
Series:Visual Intelligence
Subjects:
Online Access:https://doi.org/10.1007/s44267-024-00066-7
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1850109782623518720
author Dehong Kong
Siyuan Liang
Xiaopeng Zhu
Yuansheng Zhong
Wenqi Ren
author_facet Dehong Kong
Siyuan Liang
Xiaopeng Zhu
Yuansheng Zhong
Wenqi Ren
author_sort Dehong Kong
collection DOAJ
description Abstract Visual language pre-training (VLP) models have demonstrated significant success in various domains, but they remain vulnerable to adversarial attacks. Addressing these adversarial vulnerabilities is crucial for enhancing security in multi-modal learning. Traditionally, adversarial methods that target VLP models involve simultaneous perturbation of images and text. However, this approach faces significant challenges. First, adversarial perturbations often fail to translate effectively into real-world scenarios. Second, direct modifications to the text are conspicuously visible. To overcome these limitations, we propose a novel strategy that uses only image patches for attacks, thus preserving the integrity of the original text. Our method leverages prior knowledge from diffusion models to enhance the authenticity and naturalness of the perturbations. Moreover, to optimize patch placement and improve the effectiveness of our attacks, we utilize the cross-attention mechanism, which encapsulates inter-modal interactions by generating attention maps to guide strategic patch placement. Extensive experiments conducted in a white-box setting for image-to-text scenarios reveal that our proposed method significantly outperforms existing techniques, achieving a 100% attack success rate.
format Article
id doaj-art-c87c92f554ec49888574afaf3ebf4399
institution OA Journals
issn 2731-9008
language English
publishDate 2024-12-01
publisher Springer
record_format Article
series Visual Intelligence
spelling doaj-art-c87c92f554ec49888574afaf3ebf43992025-08-20T02:37:58ZengSpringerVisual Intelligence2731-90082024-12-012111010.1007/s44267-024-00066-7Patch is enough: naturalistic adversarial patch against vision-language pre-training modelsDehong Kong0Siyuan Liang1Xiaopeng Zhu2Yuansheng Zhong3Wenqi Ren4School of Cyber Science and TechnologySchool of Computing, National University of SingaporeGuangdong Testing Institute of Product Quality SupervisionGuangdong Testing Institute of Product Quality SupervisionSchool of Cyber Science and TechnologyAbstract Visual language pre-training (VLP) models have demonstrated significant success in various domains, but they remain vulnerable to adversarial attacks. Addressing these adversarial vulnerabilities is crucial for enhancing security in multi-modal learning. Traditionally, adversarial methods that target VLP models involve simultaneous perturbation of images and text. However, this approach faces significant challenges. First, adversarial perturbations often fail to translate effectively into real-world scenarios. Second, direct modifications to the text are conspicuously visible. To overcome these limitations, we propose a novel strategy that uses only image patches for attacks, thus preserving the integrity of the original text. Our method leverages prior knowledge from diffusion models to enhance the authenticity and naturalness of the perturbations. Moreover, to optimize patch placement and improve the effectiveness of our attacks, we utilize the cross-attention mechanism, which encapsulates inter-modal interactions by generating attention maps to guide strategic patch placement. Extensive experiments conducted in a white-box setting for image-to-text scenarios reveal that our proposed method significantly outperforms existing techniques, achieving a 100% attack success rate.https://doi.org/10.1007/s44267-024-00066-7Adversarial PatchPhysical AttackDiffusion ModelNaturalistic
spellingShingle Dehong Kong
Siyuan Liang
Xiaopeng Zhu
Yuansheng Zhong
Wenqi Ren
Patch is enough: naturalistic adversarial patch against vision-language pre-training models
Visual Intelligence
Adversarial Patch
Physical Attack
Diffusion Model
Naturalistic
title Patch is enough: naturalistic adversarial patch against vision-language pre-training models
title_full Patch is enough: naturalistic adversarial patch against vision-language pre-training models
title_fullStr Patch is enough: naturalistic adversarial patch against vision-language pre-training models
title_full_unstemmed Patch is enough: naturalistic adversarial patch against vision-language pre-training models
title_short Patch is enough: naturalistic adversarial patch against vision-language pre-training models
title_sort patch is enough naturalistic adversarial patch against vision language pre training models
topic Adversarial Patch
Physical Attack
Diffusion Model
Naturalistic
url https://doi.org/10.1007/s44267-024-00066-7
work_keys_str_mv AT dehongkong patchisenoughnaturalisticadversarialpatchagainstvisionlanguagepretrainingmodels
AT siyuanliang patchisenoughnaturalisticadversarialpatchagainstvisionlanguagepretrainingmodels
AT xiaopengzhu patchisenoughnaturalisticadversarialpatchagainstvisionlanguagepretrainingmodels
AT yuanshengzhong patchisenoughnaturalisticadversarialpatchagainstvisionlanguagepretrainingmodels
AT wenqiren patchisenoughnaturalisticadversarialpatchagainstvisionlanguagepretrainingmodels