Enhancing practicality and efficiency of deepfake detection

Area of Science:

• Computer Science
• Artificial Intelligence
• Digital Forensics

Background:

• The rapid advancement of deepfake technology poses significant challenges for content authenticity verification.
• Existing deepfake detection methods are computationally intensive, limiting their practical application, especially in resource-constrained environments like edge computing.
• There is a growing need for efficient and accessible deepfake detection solutions for non-expert users.

Purpose of the Study:

• To develop and present a series of techniques for accelerating the inference speed of deepfake detection on video data.
• To explore the application of steganalysis principles for identifying deepfakes.
• To identify key strategies for reducing the computational footprint of deepfake detection models, specifically convolutional neural networks.

Main Methods:

• Proposed novel techniques to optimize deepfake detection inference speed on video.
• Adapted steganalysis methods to treat deepfakes as encoded secret payloads within images.
• Identified critical factors for significantly downsizing core convolutional neural network architectures.
• Evaluated the proposed methods on Celeb-DFv2 and DFDC datasets.

Main Results:

• Achieved competitive performance in deepfake detection accuracy on benchmark datasets.
• Demonstrated a substantial reduction in computational cost and resource requirements compared to traditional methods.
• Validated the effectiveness of the proposed acceleration and model size reduction techniques.

Conclusions:

• The developed techniques offer a practical and efficient approach to deepfake detection, suitable for a wider range of applications, including edge computing.
• The findings suggest that deepfake detection can be made more accessible and less resource-intensive without compromising accuracy.
• This research contributes to the ongoing efforts to combat the spread of misinformation and enhance digital media integrity.