Face Manipulation Detection Based on Supervised Multi-Feature Fusion Attention Network.

Area of Science:

• Computer Science
• Artificial Intelligence
• Digital Forensics

Background:

• Deep learning advancements allow for highly realistic video face manipulation, posing risks for misinformation and malicious attacks.
• Existing face manipulation detection methods often rely on single convolutional neural networks, which may not fully capture human visual mechanisms or rich image details.
• The limitations of current methods hinder accurate detection, especially for subtle manipulations.

Purpose of the Study:

• To propose a novel and accurate face manipulation detection method.
• To address the limitations of single-feature extraction in existing methods.
• To enhance the detection of subtle and realistic face manipulations in videos.

Main Methods:

• Development of a supervised multi-feature fusion attention network (SMFAN).
• Integration of SMFAN with a capsule network for enhanced feature extraction.
• Utilization of focal loss for effective hard example mining.

Main Results:

• The proposed SMFAN method demonstrated superior performance on the FaceForensics++ dataset.
• The multi-feature fusion and attention mechanisms improved the extraction of crucial details in manipulated faces.
• The method showed enhanced accuracy in distinguishing real from fake faces compared to existing approaches.

Conclusions:

• The novel SMFAN method offers a significant improvement in face manipulation detection.
• Combining capsule networks with multi-feature fusion attention effectively captures subtle manipulation artifacts.
• This approach provides a more robust solution for identifying deepfake videos.