A Dual-Branch Fusion Model for Deepfake Detection Using Video Frames and Microexpression Features.

Area of Science:

• Computer Science
• Artificial Intelligence
• Digital Forensics

Background:

• The proliferation of synthetic media, or deepfakes, poses significant challenges due to potential malicious applications.
• Existing deepfake detection methods often struggle with sophisticated manipulations, necessitating advancements in detection accuracy and robustness.

Purpose of the Study:

• To develop and validate a novel deepfake detection model integrating spatiotemporal video features with subtle facial microexpression analysis.
• To enhance the accuracy and reliability of deepfake detection systems by leveraging complementary data modalities.

Main Methods:

• A dual-branch fusion model was designed, employing a 3D ResNet18 for extracting spatiotemporal features from video frames.
• A transformer model was utilized to meticulously capture and analyze facial microexpression patterns, which are inherently difficult to synthesize accurately.
• The combined model was rigorously evaluated on the comprehensive FaceForensics++ (FF++) dataset.

Main Results:

• The proposed deepfake detection approach achieved an exceptional accuracy rate of 99.81%.
• The model attained a perfect Receiver Operating Characteristic Area Under the Curve (ROC-AUC) score of 100%, indicating superior discrimination capabilities.
• Performance benchmarks demonstrated that the novel method surpasses current state-of-the-art deepfake detection techniques.

Conclusions:

• Integrating diverse feature sets, specifically spatiotemporal video data and microexpression dynamics, is crucial for robust deepfake detection.
• The developed dual-branch fusion model offers a significant advancement in combating the misuse of synthetic media.
• This research underscores the potential of hybrid approaches in addressing the evolving landscape of digital media manipulation.