Transferable Adversarial Face Attack with Text Controlled Attribute

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Li, Wenyun, Zhang, Zheng, Lan, Xiangyuan, Jiang, Dongmei
Formato: Preprint
Publicado: 2024
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866929693412818944
author Li, Wenyun
Zhang, Zheng
Lan, Xiangyuan
Jiang, Dongmei
author_facet Li, Wenyun
Zhang, Zheng
Lan, Xiangyuan
Jiang, Dongmei
contents Traditional adversarial attacks typically produce adversarial examples under norm-constrained conditions, whereas unrestricted adversarial examples are free-form with semantically meaningful perturbations. Current unrestricted adversarial impersonation attacks exhibit limited control over adversarial face attributes and often suffer from low transferability. In this paper, we propose a novel Text Controlled Attribute Attack (TCA$^2$) to generate photorealistic adversarial impersonation faces guided by natural language. Specifically, the category-level personal softmax vector is employed to precisely guide the impersonation attacks. Additionally, we propose both data and model augmentation strategies to achieve transferable attacks on unknown target models. Finally, a generative model, \textit{i.e}, Style-GAN, is utilized to synthesize impersonated faces with desired attributes. Extensive experiments on two high-resolution face recognition datasets validate that our TCA$^2$ method can generate natural text-guided adversarial impersonation faces with high transferability. We also evaluate our method on real-world face recognition systems, \textit{i.e}, Face++ and Aliyun, further demonstrating the practical potential of our approach.
format Preprint
id arxiv_https___arxiv_org_abs_2412_11735
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Transferable Adversarial Face Attack with Text Controlled Attribute
Li, Wenyun
Zhang, Zheng
Lan, Xiangyuan
Jiang, Dongmei
Computer Vision and Pattern Recognition
Artificial Intelligence
Traditional adversarial attacks typically produce adversarial examples under norm-constrained conditions, whereas unrestricted adversarial examples are free-form with semantically meaningful perturbations. Current unrestricted adversarial impersonation attacks exhibit limited control over adversarial face attributes and often suffer from low transferability. In this paper, we propose a novel Text Controlled Attribute Attack (TCA$^2$) to generate photorealistic adversarial impersonation faces guided by natural language. Specifically, the category-level personal softmax vector is employed to precisely guide the impersonation attacks. Additionally, we propose both data and model augmentation strategies to achieve transferable attacks on unknown target models. Finally, a generative model, \textit{i.e}, Style-GAN, is utilized to synthesize impersonated faces with desired attributes. Extensive experiments on two high-resolution face recognition datasets validate that our TCA$^2$ method can generate natural text-guided adversarial impersonation faces with high transferability. We also evaluate our method on real-world face recognition systems, \textit{i.e}, Face++ and Aliyun, further demonstrating the practical potential of our approach.
title Transferable Adversarial Face Attack with Text Controlled Attribute
topic Computer Vision and Pattern Recognition
Artificial Intelligence
url https://arxiv.org/abs/2412.11735