Transformer-based cascade networks with spatial and channel reconstruction convolution for deepfake detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• The proliferation of sophisticated forged video technology poses significant threats to individuals, society, and national security.
• The rapid advancement of deepfake technology necessitates the development of robust and up-to-date detection models.
• Training deepfake detection models requires substantial volumes of diverse data, presenting a data management challenge.

Purpose of the Study:

• To propose a novel deepfake detection model that addresses the limitations of existing methods.
• To enhance the accuracy and efficiency of deepfake detection through innovative network architecture.
• To provide a reliable solution for identifying forged videos in the digital landscape.

Main Methods:

• A cascade network architecture integrating spatial and channel reconstruction convolution (SCConv) with a vision transformer was developed.
• The model utilizes SCConv and regular convolution in its initial stages for fake video detection, followed by a vision transformer.
• The feed-forward layer of the vision transformer was enhanced to improve detection accuracy and reduce computational load.
• Datasets were processed by splitting video frames and extracting faces to create comprehensive real and fake face image datasets.

Main Results:

• The proposed model achieved high detection accuracies across multiple benchmark datasets: 87.92% on DFDC, 99.23% on FaceForensics++, and 99.98% on Celeb-DF.
• The enhanced vision transformer component successfully increased detection accuracy while simultaneously reducing the model's computational requirements.
• Qualitative assessments, including visualization results, demonstrated the model's effectiveness in authenticating video content.

Conclusions:

• The developed deepfake detection model, leveraging SCConv and an enhanced vision transformer, offers a highly effective solution for identifying forged videos.
• The model's superior performance on diverse datasets and its computational efficiency make it a valuable tool in combating the spread of malicious deepfakes.
• The study validates the model's efficacy through rigorous testing and confirms its potential for real-world application in deepfake detection.