Introduction

This repository is an implementation of the paper ForensicsForest Family: A Series of Multi-scale Hierarchical Cascade Forests for Detecting GAN-generated Faces presented in TIFS 2024. This paper describes ForensicsForest Family, a novel set of forest-based methods to detect GAN-generate faces. In contrast to the recent efforts of using CNNs, we investigate the feasibility of using forest models and introduce a series of multi-scale hierarchical cascade forests, which are ForensicsForest, Hybrid ForensicsForest, and Divide-and-Conquer ForensicsForest, respectively.

ForensicsForest

Requirements

python 3.7
numpy
pandas
deep-forest 0.1.7
dlib 19.24.0
scikit-learn 1.0.2
opencv-python
xgboost 1.6.2
torch 1.13.1
torchvision 0.14.1

Dataset Preparation

We include the dataset loaders for several commonly-used generated-faces datasets, i.e., StyleGAN , StyleGAN2 , and StyleGAN3 . You can enter the dataset website to download the original data. The folder structure is shown as following. Specially, 0 denotes fake and 1 denotes real.

StyleGAN
|---train
    |---0
    |---1
|---test
    |---0
    |---1
|---train.csv
|---test.csv

To obtain the CSV file of the dataset, run the following script in your console.

run get_csv.py
     --dataset_path /media/ForensicsForest-main/StyleGAN/train/ \
     --save_csv_path /media/ForensicsForest-main/StyleGAN/train.csv

Input Feature Extract

We use four scales as N = 1, 2, 3, 4. Then we can extract appearance and frequency features from N(N ≥ 1) patches of input images. Specially, we extract biology features from the whole image. You can extract the input features for each scale using extract_feature.py as following.(Our code uses N=4 as an example).

run extract_feature.py
    --m 2  \
    --n 2  \
    --detector_path /media/ForensicsForest-main/shape_predictor_68_face_landmarks.dat  \
    --read_path  /media/ForensicsForest-main/StyleGAN/train/0/  \
    --save_patch_path  /media/ForensicsForest-main/StyleGAN/N=4/patch/train/0/  \
    --save_feature_path1 /media/ForensicsForest-main/StyleGAN/N=4/train/0/hist/  \
    --save_feature_path2 /media/ForensicsForest-main/StyleGAN/N=4/train/0/spec/  \
    --save_feature_path3 /media/ForensicsForest-main/StyleGAN/N=4/train/0/landmarks/  \
    --save_feature_path /media/ForensicsForest-main/StyleGAN/N=4/train/0/

Note that extract_feature.py is only for the folder corresponding to the generated faces, you can modify the file path to extract the input features of the real faces, and then use merge.py to concatenate the extracted features as following.

run merge.py
    --train_feature_path /media/ForensicsForest-main/StyleGAN/N=4/train/  \
    --test_feature_path /media/ForensicsForest-main/StyleGAN/N=4/test/  \
    --save_path  /media/ForensicsForest-main/StyleGAN/N=4/

Hierarchical Cascade Forest

For the cascade.py of package deep-forest, you should replace it with ForensicsForest_cascade.py.

Multi-scale Ensemble

Run main2.py, main3.py and main4.py respectively to obtain the augmented features of each sacle as following.

run main2.py --dataset_path /media/ForensicsForest-main/StyleGAN/

run main3.py --dataset_path /media/ForensicsForest-main/StyleGAN/

run main4.py --dataset_path /media/ForensicsForest-main/StyleGAN/

Then run main1.py for final results as following.

run main1.py --dataset_path /media/ForensicsForest-main/StyleGAN/

Notice: For Hybrid ForensicsForest and Divide-and-Conquer ForensicsForest, you just need to change Module Hierarchical Cascade Forest as following.

Hybrid ForensicsForest

For the cascade.py of package deep-forest, you can replace it with Hybrid_FF_cascade.py.

Divide-and-Conquer ForensicsForest

For the cascade.py of package deep-forest, you can replace it with D-and-C_FF_cascade.py. For the layer.py of package deep-forest, you can replace it with D-and-C_FF_layer.py.

Citation

This paper is extended from our ICME conference paper(Oral) Forensics Forest: Multi-scale Hierarchical Cascade Forest for Detecting GAN-generated Faces.

If you find our ForensicsForest useful to your research, please cite it as follows:

@article{lu2024forensicsforest,
title={ForensicsForest Family: A Series of Multi-scale Hierarchical Cascade Forests for Detecting GAN-generated Faces},
author={Lu, Jiucui and Zhou, Jiaran and Li, Bin and Dong, Junyu and Lyu, Siwei and Li, Yuezun},
journal={IEEE Transactions on Information Forensics and Security (TIFS)},
year={2024}
}

@inproceedings{lu2023forensics,
  title={Forensics Forest: Multi-scale Hierarchical Cascade Forest for Detecting GAN-generated Faces},
  author={Lu, Jiucui and Li, Yuezun and Zhou, Jiaran and Li, Bin and Lyu, Siwei},
  booktitle={2023 IEEE International Conference on Multimedia and Expo (ICME)},
  pages={2309--2314},
  year={2023},
  organization={IEEE}
}