CSTAN: A Deepfake Detection Network with CST Attention for Superior Generalization.

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Deepfake technology poses significant security risks to facial recognition systems.
• Current deep learning-based deepfake detection models often lack cross-dataset generalization.
• High accuracy within datasets does not guarantee performance on unseen data.

Purpose of the Study:

• To develop a deepfake detection model with improved cross-dataset generalization.
• To enhance the model's ability to learn features from image forgery regions.
• To address the limitations of existing deepfake detection methods.

Main Methods:

• Proposed the Channel-Spatial-Triplet Attention Network (CSTAN) for deepfake detection.
• Introduced the Channel-Spatial-Triplet (CST) attention mechanism for subtle local information extraction.
• Developed OD-ResNet-34, a novel feature extraction method using ODConv for dynamic adaptability.

Main Results:

• The CSTAN model demonstrated superior generalization ability on cross-datasets (Celeb-DF-v1, Celeb-DF-v2) compared to similar models.
• The CST attention mechanism effectively captured feature channels and spatial correlations across multiple scales.
• OD-ResNet-34 enhanced the model's adaptability to diverse data features.

Conclusions:

• The proposed CSTAN model offers enhanced generality for deepfake detection across different datasets.
• The integration of CST attention and OD-ResNet-34 contributes to more robust deepfake detection.
• This research advances the development of reliable deepfake detection systems against sophisticated forgeries.