AdvFaceGAN: a face dual-identity impersonation attack method based on generative adversarial networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Current facial recognition systems have security vulnerabilities.
• Existing digital and physical attack methods have limitations in real-world scenarios.
• Digital attacks lack physical implementation, while physical attacks lack stealth.

Purpose of the Study:

• To reveal security vulnerabilities in commercial facial recognition systems.
• To promote advancements in facial recognition technology security.
• To propose a novel dual-identity impersonation attack method.

Main Methods:

• AdvFaceGAN, a generative adversarial network (GAN)-based method, generates dual-identity adversarial faces.
• Introduces structural similarity loss for enhanced visual quality and stealthiness.
• Employs an ensemble of diverse facial recognition models and incorporates source identity loss for improved attack effectiveness.

Main Results:

• AdvFaceGAN generates adversarial faces with excellent stealthiness, passing human review.
• Achieves superior traditional and dual-identity impersonation attack success rates.
• Demonstrates better stealthiness and transferability in black-box scenarios compared to existing methods.

Conclusions:

• AdvFaceGAN effectively addresses limitations in existing impersonation attack methods.
• The proposed method enhances facial recognition security by highlighting vulnerabilities.
• Advancements in adversarial face generation can improve the robustness of facial recognition systems.