CrossDF: improving cross-domain deepfake detection with deep information decomposition

Area of Science:

• Computer Science
• Artificial Intelligence
• Digital Forensics

Background:

• Deepfake technology poses significant risks to public safety and confidence.
• Current deepfake detection methods struggle with cross-dataset generalization, failing when encountering novel manipulation techniques.
• Existing approaches often focus on specific visual anomalies, limiting their robustness.

Purpose of the Study:

• To develop a robust framework for cross-dataset deepfake detection (CrossDF).
• To improve the generalization capability of deepfake detection models to unseen manipulation techniques.
• To enhance the reliability of deepfake identification across diverse datasets and methods.

Main Methods:

• Proposed a Deep Information Decomposition (DID) framework that decomposes facial representations into deepfake-relevant and unrelated components.
• Focused on high-level semantic attributes rather than low-level visual artifacts for classification.
• Introduced an adversarial mutual information minimization strategy to enhance feature separability and decorrelation learning.

Main Results:

• Achieved an AUC of 0.779 in cross-dataset evaluation from FF++ to CDF2.
• Significantly improved the state-of-the-art AUC from 0.669 to 0.802 on a diffusion-based Text-to-Image dataset.
• Demonstrated superior effectiveness and robustness of the DID framework against unseen deepfake techniques.

Conclusions:

• The proposed DID framework offers a significant advancement in cross-dataset deepfake detection.
• By focusing on semantic attributes and employing adversarial learning, the framework achieves improved generalization.
• The DID framework enhances robustness against novel manipulation techniques, contributing to more reliable media authentication.