Evaluation of deepfake detection using YOLO with local binary pattern histogram.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Forensics

Background:

• Deepfake technology poses a significant threat by creating sophisticated forged images and videos.
• Existing deepfake detection methods often struggle with novel threats and false positives.
• The need for robust video authenticity verification is critical to combat misinformation.

Purpose of the Study:

• To propose a novel deepfake detection method, YOLO-LBPH, combining You Only Look Once and Local Binary Pattern Histogram.
• To enhance the accuracy and reliability of detecting forged videos.
• To address the limitations of current image processing techniques in identifying sophisticated fake content.

Main Methods:

• Utilized You Only Look Once (YOLO) for efficient face detection in video frames.
• Employed EfficientNet-B5 for extracting spatial features from detected face images.
• Integrated Local Binary Pattern Histogram (LBPH) to extract temporal features from spatial data.

Main Results:

• Achieved a precision score of 86.88% on the CelebDF-FaceForensics++(c23) dataset.
• Demonstrated high recall rates: 92.45% on CelebDF-FaceForensics++(c23), 93.76% on DFFD, and 94.35% on CASIA-WebFace.
• The YOLO-LBPH model showed strong performance across multiple large-scale deepfake datasets.

Conclusions:

• The proposed YOLO-LBPH method offers a promising and effective approach for deepfake video detection.
• The combination of YOLO for face detection and LBPH for feature extraction significantly improves detection accuracy.
• This technique provides a valuable tool for enhancing digital forensics and combating the spread of manipulated media.