A statistical analysis for deepfake videos forgery traces recognition followed by a fine-tuned InceptionResNetV2...

Area of Science:

• Computer Science
• Artificial Intelligence
• Digital Forensics

Background:

• Deepfake technology is rapidly advancing due to AI and deep learning.
• The proliferation of deepfakes poses significant risks to propaganda, privacy, and security.
• Existing deepfake detection methods require improvement to address sophisticated manipulations.

Purpose of the Study:

• To develop an analytically novel method for detecting deepfake videos.
• To address the challenges posed by increasingly competent deepfake content.
• To enhance the accuracy and reliability of deepfake identification.

Main Methods:

• Analyzing temporal discrepancies in statistical video features at the pixel level.
• Integrating spatial information from individual frames and temporal correlations between frames.
• Developing a novel Euclidean distance variation probability score for authenticity assessment.
• Fine-tuning the InceptionResNetV2 model with an added dense layer for deepfake detection using the FaceForensics++ dataset.

Main Results:

• The proposed method achieved 99.80% accuracy on the FaceForensics++ (FF++) dataset.
• The method demonstrated 97.60% accuracy on the CelebDF dataset.
• The fine-tuned InceptionResNetV2 model outperformed existing deepfake detection techniques on unseen data.

Conclusions:

• The novel method effectively detects deepfake videos by analyzing subtle temporal and spatial aberrations.
• The proposed approach offers a significant advancement in deepfake detection accuracy and reliability.
• This research contributes a robust solution to mitigate the risks associated with deepfake technology.