<p>Concealed object detection (COD) has advanced significantly and is crucial in various fields. However, it raises new security and privacy issues, as powerful COD models can potentially reveal sensitive information like human privacy organs or military camouflage. In this paper, we address this issue through the lens of adversarial attacks and introduce a new task: Adversarial Attacks against COD. Compared to general adversarial attacks on object detection models, this new task presents an additional challenge. The challenge lies in generating adversarial perturbations that disrupt the differential information contained within various scenes simultaneously. To address this, In this paper, we introduce a novel adversarial attack method, context-aware target texture perturbation (<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="371_2025_3805_Article_IEq1.gif" Format="GIF" Height="17" Rendition="HTML" Resolution="72" Type="Linedraw" Width="42" /> </InlineMediaObject> <EquationSource Format="TEX">$\hbox {CAT}^2$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mtext>CAT</mtext> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation>P), specifically designed to fool COD models. <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="371_2025_3805_Article_IEq1.gif" Format="GIF" Height="17" Rendition="HTML" Resolution="72" Type="Linedraw" Width="42" /> </InlineMediaObject> <EquationSource Format="TEX">$\hbox {CAT}^2$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mtext>CAT</mtext> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation>P generates adversarial perturbations based on background texture information, disrupting the differential features used by COD models to distinguish concealed objects. The attack comprises three modules: perturbation generation, target localization, and perturbation bootstrap. Extensive experiments on benchmark datasets demonstrate <InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="371_2025_3805_Article_IEq1.gif" Format="GIF" Height="17" Rendition="HTML" Resolution="72" Type="Linedraw" Width="42" /> </InlineMediaObject> <EquationSource Format="TEX">$\hbox {CAT}^2$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mtext>CAT</mtext> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation>P’s effectiveness in reducing COD model performance by up to 40% while preserving the visual quality of original images. This work highlights the security vulnerabilities of COD models and provides insights into evaluating their robustness.</p>

Context-aware target texture perturbation attack for concealed object detection

期刊论文
发表日期：

2025年1月23日

查看全文

Jialin Zhang,
Xiao Wang,
Hui Wei,
Kui Jiang,
Nan Mu,
Zheng Wang

摘要

Concealed object detection (COD) has advanced significantly and is crucial in various fields. However, it raises new security and privacy issues, as powerful COD models can potentially reveal sensitive information like human privacy organs or military camouflage. In this paper, we address this issue through the lens of adversarial attacks and introduce a new task: Adversarial Attacks against COD. Compared to general adversarial attacks on object detection models, this new task presents an additional challenge. The challenge lies in generating adversarial perturbations that disrupt the differential information contained within various scenes simultaneously. To address this, In this paper, we introduce a novel adversarial attack method, context-aware target texture perturbation ( $\hbox {CAT}^2$ ${CAT}^{2}$ P), specifically designed to fool COD models. $\hbox {CAT}^2$ ${CAT}^{2}$ P generates adversarial perturbations based on background texture information, disrupting the differential features used by COD models to distinguish concealed objects. The attack comprises three modules: perturbation generation, target localization, and perturbation bootstrap. Extensive experiments on benchmark datasets demonstrate $\hbox {CAT}^2$ ${CAT}^{2}$ P’s effectiveness in reducing COD model performance by up to 40% while preserving the visual quality of original images. This work highlights the security vulnerabilities of COD models and provides insights into evaluating their robustness.