Adversarial Patch Attacks on Deep-Learning-Based Face Recognition Systems Using Generative Adversarial Networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning is widely used in face recognition for security and access control.
• Deep learning models are vulnerable to adversarial attacks, which can manipulate images or use adversarial patches.
• Most adversarial attack research assumes white-box access, which is unrealistic for real-world scenarios.

Purpose of the Study:

• To propose a novel method for generating adversarial patches against face recognition systems.
• To evaluate the effectiveness of adversarial patch attacks in a black-box setting.
• To demonstrate a higher attack success rate compared to existing adversarial attack methods.

Main Methods:

• A Generative Adversarial Network (GAN) was employed to create adversarial patches.
• The proposed method targets face recognition systems treated as black boxes, with no prior knowledge of their architecture or parameters.
• Dodging and impersonation attacks were simulated using the generated adversarial patches.

Main Results:

• The proposed Generative Adversarial Network method successfully generated adversarial patches.
• The adversarial patch attacks demonstrated a significant success rate in both dodging and impersonation scenarios.
• The method achieved a higher attack success rate compared to previous adversarial attack techniques.

Conclusions:

• Adversarial patch attacks on face recognition systems are feasible and effective, even in black-box scenarios.
• Generative Adversarial Networks offer a powerful tool for creating sophisticated adversarial examples.
• Further research is needed to develop robust defenses against such black-box adversarial attacks.