Revisiting Face Forgery Detection: From Facial Representation to Forgery Detection

Area of Science:

• Computer Science
• Artificial Intelligence
• Image Processing

Background:

• Face Forgery Detection (FFD) models struggle with generalization due to overfitting on specific forgery patterns from diverse synthesis algorithms.
• Current methods using general pre-trained backbones lack domain-specific facial knowledge, hindering the identification of subtle forgery cues.

Purpose of the Study:

• To enhance the generalization capabilities of Face Forgery Detection (FFD) models.
• To develop a more effective FFD workflow by integrating domain-specific facial representation and forgery detection.
• To improve the identification of implicit forgery cues and inference reliability.

Main Methods:

• Developed an FFD-specific backbone through self-supervised pre-training on real faces for superior facial representation.
• Proposed a competitive fine-tuning framework to stimulate the backbone in identifying implicit forgery cues.
• Devised a threshold optimization mechanism using prediction confidence to enhance inference reliability.

Main Results:

• The proposed method achieved excellent performance in Face Forgery Detection (FFD).
• Demonstrated strong generalization capabilities on unseen forgery patterns.
• Showcased effectiveness in related face-related tasks, including presentation attack detection.

Conclusions:

• The FFD-specific pre-trained backbone and competitive fine-tuning framework significantly improve generalization and performance in deepfake detection.
• The integration of domain-specific knowledge and advanced training techniques is crucial for robust face forgery detection.
• The method offers improved reliability and effectiveness for identifying digital face forgeries.