Deepfake video deception detection using visual attention-based method.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Forensics

Background:

• The proliferation of artificial digital data raises concerns about authenticity.
• Deepfake technology, utilizing computer vision, poses a significant threat by enabling identity concealment.
• There is a critical need for reliable methods to verify the legitimacy of facial images and videos.

Purpose of the Study:

• To develop a deep learning model capable of differentiating between real and deepfake visual content.
• To enhance the security and trustworthiness of digital media.
• To address the societal anxiety caused by the misuse of deepfake technology.

Main Methods:

• A deep learning model incorporating a visual attention strategy was developed.
• Facial regions were extracted from video frames and processed using a pre-trained ResNeXt-50 Convolutional Neural Network (CNN).
• A visual attention mechanism was employed to detect deepfake-specific artifacts.

Main Results:

• The proposed model demonstrated superior performance in identifying deepfake content.
• Evaluation was conducted under cross-dataset conditions, using Face Forensic++ C23 for training.
• The model achieved strong independent testing results on Celeb-DFv2 and DFDC datasets.

Conclusions:

• The developed deep learning model with visual attention is effective in detecting deepfake videos.
• This approach offers a robust solution for verifying the authenticity of facial media.
• The findings contribute to combating the malicious use of deepfake technology and securing digital content.