Learning Patch-Channel Correspondence for Interpretable Face Forgery Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Forensics

Background:

• High accuracy is crucial for face forgery detection models.
• Interpretability is essential for deploying these models in real-world visual content analysis.
• Current methods often lack transparency, making it difficult to understand their decision-making process.

Purpose of the Study:

• To develop an interpretable face forgery detection model.
• To learn patch-channel correspondence for enhanced feature representation.
• To provide visualized evidence for forgery detection while maintaining high accuracy.

Main Methods:

• Proposed a novel approach using patch-channel correspondence.
• Embedded a feature reorganization layer into a deep neural network.
• Utilized alternate optimization for classification and correspondence tasks.
• Implemented channel-wise decorrelation and patch-channel alignment for interpretable features.

Main Results:

• The model learns multi-channel interpretable features, with each channel corresponding to a facial patch.
• Achieved high accuracy in face forgery detection.
• Provided discriminative localization of forgery evidence through visualized features.
• Demonstrated effectiveness on popular benchmarks.

Conclusions:

• The proposed patch-channel correspondence method enhances interpretability in face forgery detection.
• The approach successfully balances high detection accuracy with clear visualization of forgery evidence.
• This work contributes to more trustworthy AI systems in visual content analysis.