3D-Aware Adversarial Makeup Generation for Facial Privacy Protection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Cybersecurity

Background:

• Social media poses significant privacy risks to facial data, making it vulnerable to unauthorized access and identification.
• Current methods for protecting facial data, such as adversarial examples, often result in poor image quality and limited transferability, hindering real-world applications.
• Existing facial recognition (FR) systems lack robust defenses against sophisticated privacy attacks.

Purpose of the Study:

• To develop a novel method for generating high-quality and transferable adversarial makeup to conceal identity information.
• To improve the robustness of facial data privacy against malicious facial recognition systems.
• To address the limitations of existing adversarial examples in terms of image quality and transferability.

Main Methods:

• Proposed a 3D-Aware Adversarial Makeup Generation Generative Adversarial Network (3DAM-GAN).
• Designed a UV-based generator with a Makeup Adjustment Module (MAM) and Makeup Transfer Module (MTM) leveraging facial symmetry.
• Implemented a makeup attack mechanism with an ensemble training strategy to enhance transferability against black-box models.

Main Results:

• 3DAM-GAN generates realistic and robust synthetic makeup for identity concealment.
• The proposed method significantly improves the quality and transferability of adversarial examples.
• Demonstrated effective protection against state-of-the-art and commercial facial recognition models and APIs.

Conclusions:

• 3DAM-GAN offers a promising solution for enhancing facial data privacy on social media.
• The method provides a practical approach to protect identities from unauthorized facial recognition.
• The improved transferability and quality of adversarial makeup open new avenues for privacy-preserving AI applications.