ID-Guard: A Universal Framework for Combating Facial Manipulation via Breaking Identification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Deep learning-based facial manipulation poses a significant threat to civil rights.
• Current defense methods may leave identifiable facial features intact, risking individual stigmatization.
• Proactive defenses embed adversarial perturbations to disrupt manipulation.

Purpose of the Study:

• To propose a universal framework, ID-Guard, for combating facial manipulation.
• To develop a method that suppresses identifiable features in manipulated faces.
• To enhance cross-model performance and defense against diverse manipulation types.

Main Methods:

• ID-Guard employs an encoder-decoder network for cross-model transferable adversarial perturbations.
• A novel Identity Destruction Module (IDM) is introduced to suppress identifiable features.
• Perturbation generation is optimized via multi-task learning with dynamic weighting.

Main Results:

• ID-Guard effectively defends against various facial manipulation models.
• The framework degrades identifiable regions in manipulated images.
• Disrupted images evade facial inpainting and facial recognition systems.

Conclusions:

• ID-Guard offers a robust solution against sophisticated facial manipulation.
• The framework protects individuals by removing identifying features from manipulated images.
• ID-Guard is versatile and can be integrated into other AI security tasks.