Universal Adversarial Spoofing Attacks against Face Recognition

Amada, Takuma; Liew, Seng Pei; Kakizaki, Kazuya; Araki, Toshinori

doi:10.1109/ijcb52358.2021.9484380

Cited by 4 publications

(4 citation statements)

References 17 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…To make our attack practical in the real world, we try to improve the robustness of the generated adversarial patch in this subsection. Specifically, to mitigate the negative impact of color shift and shape deformation caused by the printingcapturing process, we have adopted two methods: (1) We have developed an MLP model that maps the color transformation from the digital world to the physical world, which is used for pre-calibrating the color shift. (2) We have incorporated the Expectation of Transformation (EoT) technique to enhance the patch's resilience against deformation.…”

Section: Physical Implementationmentioning

confidence: 99%

“…Metric. In the real-world evaluation, we use two metrics: (1) The attack success rate (ASR) defined in simulation evaluation, and (2) The attack success rate in consecutive frames (F succ ) which indicates the attack ability against the target systems in the real-world setting.…”

Section: ) Black-box Transferabilitymentioning

confidence: 99%

“…In this section, we conduct a user study to measure users' attitudes towards the enrollment phase of the face authentication systems. Our aim is to investigate (1) what types of enrollment methods are widely used (Q2 in Tab. VI), and (2) whether the enrollment is supervised in real-world scenarios (Q3 in Tab.…”

Section: User Studymentioning

confidence: 99%

“…An Universal adversarial attack is the attack strategy that allows for targeting multiple objectives using a single perturbation or patch. To against the face recognition, Amada T et al [1] utilized global pixel-level adversarial perturbations to launch the multiple-identity attack against four face recognition models, both in white-box and black-box setting. However, the practical deployment of pixellevel perturbations in the physical world remains challenging, rendering them ineffective for attacking real-world face authentication systems.…”

Section: A Multiple-identity Face Spoofing Attackmentioning

confidence: 99%

See 3 more Smart Citations

UniID: Spoofing Face Authentication System by Universal Identity

Wu,

Cheng,

Zhang

et al. 2024

Proceedings 2024 Network and Distributed System Security Symposium

View full text Add to dashboard Cite

Face authentication systems are widely employed in access control systems to ensure the security of confidential facilities. Recent works have demonstrated their vulnerabilities to adversarial attacks. However, such attacks typically require adversaries to wear disguises such as glasses or hats during every authentication, which may raise suspicion and reduce their attack impacts. In this paper, we propose the UniID attack, which allows multiple adversaries to perform face spoofing attacks without any additional disguise by enabling an insider to register a universal identity into the face authentication database by wearing an adversarial patch. To achieve it, we first select appropriate adversaries through feature engineering, then generate the desired adversarial patch with a multi-target joint-optimization approach, and finally overcome practical challenges such as improving the transferability of the adversarial patch towards black-box systems and enhancing its robustness in the physical world. We implement UniID in laboratory setups and evaluate its effectiveness with six face recognition models (FaceNet, Mobile-FaceNet, ArcFace-18/50, and MagFace-18/50) and two commercial face authentication systems (ArcSoft and Face++). Simulation and real-world experimental results demonstrate that UniID can achieve a max attack success rate of 100% and 79% in 3-user scenarios under the white-box setting and black-box setting respectively, and it can be extended to more than 8 users.

show abstract

Section: Physical Implementationmentioning

confidence: 99%

Section: ) Black-box Transferabilitymentioning

confidence: 99%

Section: User Studymentioning

confidence: 99%

Section: A Multiple-identity Face Spoofing Attackmentioning

confidence: 99%

See 2 more Smart Citations