This repository is a python library, which incorperates existing deepfake detection method into an unified framework.

To date, this toolbox supports following methods:

1. The original method (with slight modification for intergration) can be downloaded from here(From console you can gdown package). Unzip the method to folder externals. You may need to make a new folder following below structure.

   Root --- deepfor ----- externals
         |- demo.py    |- methods
         |- ...        |- ...
         
   

2. demo.py is a demo script that shows how to run different methods. In our implementation, each method corresponds to a class, eg, ClassNSeg().

   import deepfor, cv2
   
   # Read image
   img = cv2.imread('test_img.png')
   conf = deepfor.ClassNSeg().run(img) # conf of fake
   print('Fake confidence score is: {}'.format(conf))

   The class name of each method can be found in demo.py

3. Since different methods usually need different environment, we provides docker images for each method to save time. Following is an example of setting up environment.

   # Build docker enironment from docker image, eg,
   docker build -t classnseg ./dockerfile/ClassNSeg/

   Then run the demo code:

   docker run --runtime=nvidia -e NVIDIA_VISIBLE_DEVICES=0 -v /home/.../deepfake-o-meter:/deepfake-o-meter/ classnseg python deepfake-o-meter/demo.py --img=/deepfake-o-meter/test_img.png  --name=ClassNSeg

   or

   docker run --gpus all -e NVIDIA_VISIBLE_DEVICES=0 -v /home/.../deepfake-o-meter:/deepfake-o-meter/ classnseg python deepfake-o-meter/demo.py --img=/deepfake-o-meter/test_img.png  --name=ClassNSeg

   Note the first time to install the dock environment may take a while, but it can be reused since then.

• [4/13/2020] The first stage is all set, where seven methods are included.

We will employ docker to handle platform issue of each method.

Please cite our paper in your publications if this toolbox is used in your research:

@inproceedings{Celeb_DF_cvpr20,
   author = {Yuezun Li, Xin Yang, Pu Sun, Honggang Qi and Siwei Lyu},
   title = {Celeb-DF: A Large-scale Challenging Dataset for DeepFake Forensics},
   booktitle= {IEEE Conference on Computer Vision and Patten Recognition (CVPR)},
   year = {2020}
}

If you are interested in incorperating your method into this toolbox, please provide interface functions in a python file as follows:

# For instance, the interface file is called utils.py, which should contain four functions

def init_model(*args); # This is for initializing the network
def crop_face(*args); # Crop the faces from input image
def preproc(*args); # Any pre-processing operation on cropped faces, eg, normalization
def predict(*args); # Predictions of your method by taking the processed face as input

With these interfaces, we can intergrate your method more conviniently.

We also provide a GUI interface for users. The detail of usage procedure is shown here.

• Yuezun Li
• Cong Zhang
• Pu Sun
• Ming-ching Chang
• Siwei Lyu