Advanced framework for multilevel detection of digital video forgeries.

Area of Science:

• Computer Vision
• Digital Forensics
• Artificial Intelligence

Background:

• Digital media's growth fuels concerns over forged video dissemination and misuse.
• Current forgery detection methods struggle with complex, layered (multilevel) forgeries.

Purpose of the Study:

• To develop an innovative framework for detecting sophisticated two- and three-level video forgeries.
• To address the limitations of existing technologies in identifying complex forged content.

Main Methods:

• Utilized attention-augmented convolutional neural networks (AACNNs) for intricate feature extraction from forged frames.
• Employed a U-Net-based CycleGAN for accurate localization of forged regions.
• Integrated model-agnostic meta-learning to enhance detection robustness and accuracy.
• Developed and utilized a custom dataset representing complex forgery scenarios.

Main Results:

• The AACNN framework achieved 98.2% accuracy in a 10-shot scenario.
• Demonstrated high performance with 96.3% sensitivity, 97.6% specificity, and 96.8% F1-score.
• Successfully identified both two- and three-level forgeries through local and global attention mechanisms.

Conclusions:

• The proposed framework significantly advances the accuracy and reliability of sophisticated video forgery detection.
• The integration of AACNNs and CycleGAN offers a robust solution for complex digital forensics challenges.
• This research provides a critical tool for combating the misuse of advanced forged video content.