Face Real/Fake Detector

A program to determine whether a face in an image is real or AI-generated (deepfake) using gradient and covariance matrix analysis.

Method

The program implements the method described in task.md:

1. Luminance Conversion: The image is converted to grayscale (luminance)
2. Gradient Computation: Horizontal (Gx) and vertical (Gy) brightness gradients are computed
3. Covariance Matrix: A matrix M is created from all gradient vectors and the covariance matrix C = (1/N) * M^T * M is computed
4. Feature Extraction: Statistical features are extracted from the covariance matrix (eigenvalues, trace, determinant, etc.)
5. Classification: Machine learning or heuristic methods are used to determine if the image is real

Installation

pip install -r requirements.txt

Usage

Basic Usage (without training)

python face_real_fake_detector.py path/to/image.jpg

Analyze Entire Image (without face detection)

python face_real_fake_detector.py path/to/image.jpg --no-face-detection

Train Model on Labeled Data

For better accuracy, you can train the model on a dataset:

python face_real_fake_detector.py dummy.jpg --train --real-dir ./real_faces --fake-dir ./fake_faces

After training, the model will use the trained classifier for more accurate predictions.

Project Structure

• face_real_fake_detector.py - main script with FaceRealFakeDetector class
• requirements.txt - Python dependencies
• task.md - method description
• image.png - method diagram
• GRADIENTS.md - detailed explanation of gradient analysis

How It Works

1. Face Detection: Haar Cascade is used to detect faces in the image
2. Gradient Analysis: Real images have more structured and consistent gradients that correspond to the physical structure of objects
3. Statistical Analysis: The covariance matrix of gradients reveals patterns that differ between real and generated images
4. Classification: Extracted features are used to determine whether the image is real or fake

Notes

• Without training, the model uses a heuristic method with low confidence
• For best results, it is recommended to train the model on labeled data
• The method works best on face images with good lighting and resolution