TSFF-Net: A deep fake video detection model based on two-stream feature domain fusion

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Forensics

Background:

• Deepfake generation techniques, especially Generative Adversarial Networks (GANs), pose significant challenges for accurate face forgery detection.
• Existing methods struggle with complex scenes and low-quality, compressed images, limiting their real-world applicability.

Purpose of the Study:

• To develop a novel deep learning model for detecting deepfake face videos, particularly focusing on improving performance with low-quality and compressed content.
• To enhance the robustness of deepfake detection against sophisticated forgery techniques.

Main Methods:

• Proposed the Two-Stream Feature Domain Fusion Network (TSFF-Net) for deep face forgery video detection.
• Implemented spatial and frequency domain feature extraction branches, incorporating the Scharr operator for edge feature extraction.
• Integrated frequency domain information and utilized a Transformer layer for enhanced feature processing.

Main Results:

• Achieved high detection accuracies on the FaceForensics++ dataset: 97.7% (Deepfake), 91.0% (Face2Face), 98.9% (FaceSwap), and 90.0% (NeuralTextures).
• Demonstrated superior performance in detecting low-quality deepfake videos compared to existing methods.
• Showcased promising results in cross-dataset experiments, indicating generalizability.

Conclusions:

• TSFF-Net effectively addresses the limitations of current deepfake detection methods, especially for challenging low-quality scenarios.
• The fusion of spatial and frequency domain features provides a robust approach for identifying sophisticated deepfakes.
• The model's performance suggests its potential for practical applications in digital forensics and media authentication.