An efficient deepfake video detection using robust deep learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Digital Forensics

Background:

• The proliferation of synthetic media, or deepfakes, poses significant societal risks.
• Current deepfake detection methods lack universality and struggle with real-time forensic analysis.
• Identifying deepfakes is a critical first step in mitigating their spread on social media.

Purpose of the Study:

• To develop a novel, accurate, and robust deep-learning technique for identifying deepfake videos and images.
• To address the limitations of existing methods by proposing a hybrid approach for real-time forensic analysis.
• To establish a consistent basis for detecting various forms of manipulated media.

Main Methods:

• A hybrid deep-learning technique utilizing successive targeted video frames as input.
• Implementation of a ResNet-Swish-BiLSTM (Residual Network with Swish activation and Bidirectional Long Short-Term Memory) for training and classification.
• Validation using the Face Forensics (FF++) and Deepfake Detection Challenge (DFDC) datasets.

Main Results:

• Achieved 96.23% accuracy on the FF++ dataset.
• Attained 78.33% accuracy on aggregated FF++ and DFDC datasets.
• The proposed hybrid method demonstrated superior performance compared to existing techniques in identifying deepfake artifacts.

Conclusions:

• The developed ResNet-Swish-BiLSTM model offers a significant advancement in deepfake detection.
• The hybrid approach shows promise for real-time forensic applications in combating synthetic media.
• Further validation on combined datasets is recommended for broader applicability.