Multi-feature fusion based face forgery detection with local and global characteristics.

Area of Science:

• Computer Science
• Artificial Intelligence
• Information Security

Background:

• Deepfake videos pose significant societal and information security risks due to their sophisticated generation using deep learning.
• Existing deepfake detection methods often rely on single feature domains, limiting their practical effectiveness.
• The inability of the human eye to detect deepfakes necessitates advanced automated detection techniques.

Purpose of the Study:

• To develop an accurate and efficient deepfake video detection method.
• To overcome the limitations of single-feature-domain detection approaches.
• To create a robust method capable of comprehensively analyzing forgery face features.

Main Methods:

• Proposing a deepfake detection method that integrates features from the spatial, noise, and frequency domains.
• Utilizing an Inception Transformer architecture to dynamically learn the mix of global and local information.
• Evaluating the method on benchmark datasets: DFDC, Celeb-DF, and FaceForensic++.

Main Results:

• The proposed method demonstrates effectiveness and good generalization across multiple benchmark datasets.
• Achieves competitive performance compared to optimal models, despite using fewer parameters and no pre-training, distillation, or assembly.
• Ablation experiments confirm the significant contribution of each component to the overall performance.

Conclusions:

• The integrated multi-domain feature analysis approach offers a superior deepfake detection strategy.
• The Inception Transformer effectively captures complex feature interactions for robust detection.
• The method presents a promising, efficient, and generalizable solution for deepfake video identification.