Diffusion-Generated Fake Face Detection by Exploring Wavelet Domain Forgery Clues

This is the official repository of "Diffusion-Generated Fake Face Detection by Exploring Wavelet Domain Forgery Clues" (WCSP 2023)

Dependencies

• Pytorch ≥ 1.10 with CUDA ≥ 11.3
• tensorboard
• opencv-python
• timm
• pyyaml
• tqdm

Dataset and Pretrained Models

Dropbox 北航云盘 (提取码：wcsp)

• Dataset: Please unzip the downloaded datasets and put them in one folder, for example:

- /mnt/data/diffusion_detection
    - celeba_hq_256
        - test
        - train
        - val
    - wavediff_celeba256
        - ...
    - ddpm_celeba256_1000
        - ...

• Weights: put them in ./weights folder.

Start

• Training

python -W ignore main.py --folder /mnt/data/diffusion_detection --cuda 0 --batch-size 64 --save --mask True --fake ddpm_celeba256_1000 --weights ours_ddpm.pth.tar --config config/test.yaml

• Testing

python run_robust.py

parameters to modify:

• --fake To modify fake dataset;
• --cuda To modify gpu id e.g. --cuda 0 for one gpu, --cuda 0 1 for two gpus;
• --weights To modify the saving weights file name during training and testing.