MFF-Net: Deepfake Detection Network Based on Multi-Feature Fusion.

Area of Science:

• Computer Vision
• Digital Forensics
• Signal Processing

Background:

• Deepfake technology has advanced significantly, leading to widespread dissemination of forged videos and societal concerns.
• Existing deepfake detection methods often rely on RGB images, overlooking valuable spectral information.
• Textural features are commonly used but can be further enhanced by integrating signal processing techniques.

Purpose of the Study:

• To propose a novel deepfake detection network, MFF-Net, that integrates RGB features with advanced spectral textural analysis.
• To enhance the accuracy and generalization capabilities of deepfake detection systems.
• To address the limitations of current texture-based methods by incorporating signal processing.

Main Methods:

• Developed MFF-Net, a deepfake detection network incorporating Gabor convolution and residual attention blocks for feature extraction.
• Implemented a texture enhancement module to capture subtle textural details in shallow network layers.
• Utilized an attention module to focus on manipulated regions and employed feature fusion techniques.
• Introduced a diversity loss to encourage the extraction of multi-scale and multi-directional features.

Main Results:

• MFF-Net demonstrated excellent generalization across various deepfake datasets.
• The proposed method achieved state-of-the-art performance in deepfake detection.
• Fusion of RGB, spectral, and enhanced textural features proved effective.

Conclusions:

• MFF-Net offers a robust and effective approach to deepfake detection by synergistically combining different feature modalities.
• The integration of spectral analysis and advanced neural network components significantly improves detection accuracy.
• The findings contribute to the advancement of automatic deepfake detection technology, mitigating societal risks.