Research-Deepfake

This repository accompanies the paper: Fractal Characterization of Low-Correlation Signals in AI-Generated Image Detection

Source code : https://github.com/jim-xie-cn/Research-Deepfake

⚙️ Download source code

git clone https://github.com/jim-xie-cn/Research-Deepfake.git

Source Code Overview

• face_resize.py — Resize raw images to 256×256 pixels.
• face_crop.py — Detect and crop face regions from raw images.
• feature.py — Extract FD, MFS, LAC, entropy, mean, and standard deviation features.
• analyse.py — Sample feature sets for distribution analysis.

⚙️ Installation

Install required dependencies:

cd install
pip install -r requirements.txt

🔁 Reproduce Paper's Results

All experimental data of this paper is available in https://huggingface.co/datasets/jim-xie-cn/Research-Deepfake. The provided dataset already includes preprocessed images and extracted features used in the paper, enabling direct reproduction without running the full pipeline.

To reproduce the results reported in the paper follow steps:

1. Download the processed dataset

https://huggingface.co/datasets/jim-xie-cn/Research-Deepfake/resolve/main/paper/data.tar.gz?download=true

2. Extract the dataset and replace the data directory:

tar -xzvf data.tar.gz

3. Launch Jupyter Notebook:

cd scripts
jupyter notebook

4. Check result

Open the src/paper.ipynb with jupyter

Work with data of yourself

1. Dataset Preparation

Here is an example, (use 1-million-fake-faces for fake images,flickrfaceshq-dataset-ffhq for real images)

Download the following datasets and unzip them into ./data/raw:

• Fake images: 1-million-fake-faces
• Real images: flickrfaceshq-dataset-ffhq

Create the directory structure (if needed):

mkdir -p data/raw/{1-million-fake-faces,flickrfaceshq-dataset-ffhq}

After downloading and unzipping, the folders should look like:

src/
   |-face_resize.py (resize image to the same size) 
   |-face_crop.py (crop faces from images)
   |-feature.py (extract common,MFS and Lacunarity features)
   |-analyse.py (samplling and calculate statistical for common,MFS and Lacunarity features)
   |-main.ipynb (collect all data and demo)
data/
    ├── face
    │   ├── crop (cropped face images, genreated by face_crop.py)
    │   │   ├── fake
    │   │   └── real
    │   ├── features (common,MFS and Lacunarity features, genreated by feature.py)
    │   │   └── train
    |   |       |── common-10000.csv (merged common features for 10000 samples)
    |   |       |── common-50000.csv (merged common features for 50000 samples)
    │   │       ├── fake (MFS and Lacunarity features for fake images)
    │   │       └── real (MFS and Lacunarity features for real images)
    │   ├── resize (resized images, genreated by face_resize.py)
    │   │   └── 256
    │   │       ├── fake
    │   │       └── real
    │   └── status (statistical information, generated by analyse.py)
    │       └── train
    └── raw (raw images from dataset 1-million-fake-faces for fake images , dataset FFHQ for real images)
        └── dataset
            ├── 1-million-fake-faces
            └── flickrfaceshq-dataset-ffhq

2. Run process task

cd ./scripts
sh pipeline.sh

This script runs the full preprocessing and analysis pipeline end-to-end.

3. Check result

cd scripts
jupyter notebook

Open src/main.ipynb with jupyter

📚 Citation

@article{fractal_fake_detection_2026,
  title   = {Fractal Characterization of Low-Correlation Signals in AI-Generated Image Detection},
  author  = {Jim, Xie and others},
  journal = {arXiv preprint arXiv:2604.17268},
  year    = {2026}
}