Evaluating Features and Variations in Deepfake Videos Using the CoAtNet Model.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Digital Security

Background:

• Deepfake video detection is crucial for combating misinformation and enhancing digital security.
• The generalization ability of advanced AI models across diverse datasets is not fully understood.
• CoAtNet, a hybrid convolution-transformer architecture, has shown promise in computer vision tasks.

Purpose of the Study:

• To evaluate the generalization capabilities of the CoAtNet model for deepfake video detection across various datasets.
• To explore CoAtNet's performance in cross-dataset scenarios, identifying key features and variations in deepfake videos.
• To benchmark CoAtNet against state-of-the-art models in both intra-dataset and cross-dataset deepfake detection.

Main Methods:

• Extensive experiments were conducted using the CoAtNet model.
• The model was trained with diverse input and processing configurations.
• Performance was evaluated on recognized public deepfake datasets, including Celeb-DF and DFDC.

Main Results:

• CoAtNet achieved superior intra-dataset performance with an Area Under the Curve (AUC) ranging from 81.4% to 99.9%.
• The model demonstrated strong cross-dataset generalization, achieving an AUC of 78%.
• CoAtNet exhibited the best AUC for both intra-dataset and cross-dataset deepfake detection, particularly on Celeb-DF.

Conclusions:

• CoAtNet exhibits excellent generalization capabilities for deepfake video detection.
• The model's hybrid architecture effectively identifies deepfakes across different datasets.
• CoAtNet represents a significant advancement in robust deepfake detection technology.