Fully Unsupervised Deepfake Video Detection Via Enhanced Contrastive Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Deepfake videos pose significant threats to public trust and social security.
• Existing deepfake detectors often rely on supervised learning, requiring large labeled datasets.
• Supervised methods are vulnerable to insufficient or maliciously poisoned training data.

Purpose of the Study:

• To develop a fully unsupervised deepfake detection method.
• To address the limitations of supervised deepfake detection concerning data availability and integrity.
• To create a reliable deepfake detector that operates without any prior knowledge of sample labels.

Main Methods:

• A novel pseudo-label generator utilizing hand-crafted features to label training samples.
• An enhanced contrastive learner to extract and refine discriminative features iteratively.
• Binary classification based on inter-frame correlation for final deepfake detection.

Main Results:

• The proposed unsupervised detector demonstrates effectiveness across multiple benchmark datasets (FF++, Celeb-DF, DFD, DFDC, UADFV).
• Achieves performance comparable to supervised methods and superior to existing unsupervised methods.
• Shows significant superiority when dealing with poisoned or insufficient labeled training data.

Conclusions:

• The developed unsupervised deepfake detector offers a robust solution to the challenges of data scarcity and adversarial attacks.
• This approach enhances the reliability and applicability of deepfake detection in real-world scenarios.
• The method provides a powerful alternative to supervised techniques, particularly in security-sensitive contexts.