DePixel - Image Authenticity Detection

Distinguish between real and AI-generated images.

The method uses Learned Noise Patterns (LNP) and one-class classification to map real images to a dense subspace, allowing detection of generated images as outliers. This approach achieves good detection accuracy while using much less training data compared to previous methods.

Features

• Denoising network for LNP extraction
• Feature extraction from LNP amplitude spectra
• One-class SVM classifier for image authenticity prediction
• Robustness to various image post-processing operations

Installation

1 + 2. Clone the repository and create a virtual environment (optional but recommended):

git clone https://github.com/Linear-Fox-Labs/DePixel
cd DePixel
pip install -r requirements.txt

Customization

• Adjust IMAGE_SIZE in main.py to change the input image size
• Modify PLOT_FEATURES in main.py to change the number of features displayed in distribution plots
• Fine-tune classifier parameters in src/classifier.py for better performance
• Modify the config.yaml file to adjust settings.

Training

1. Prepare your dataset:

• Place real images in the data/real_images/ directory
• Place test images (real or AI-generated) in the data/test_images/ directory

2. To train the denoising network: python training/train.py

This will train the model using the real images and save the trained model as denoising_network.pth.

Output

• Console output with processing details and classification results
• Feature distribution plots saved as PNG files in the project directory

Real Image Dog1.jpg	Ai-Generated Image Dog2.jpg

Acknowledgments

• DePixel based on: Xiuli Bi, Bo Liu, et al. for their research paper "Detecting Generated Images by Real Images Only" https://arxiv.org/abs/2311.00962

• DePixel Authors: Nathan Fargo at Linear Fox Labs.