Domain General Face Forgery Detection by Learning to Weight

This code belongs to the "Domain General Face Forgery Detection by Learning to Weight"

Installation

Environment

Before run the code, you should run

$ pip install -r requirements.txt

first.

Dataset

This code is mainly focus on the GCD benchamrks, which contain three deepfake dataset: Faceforensics++,CELEB-DF,DFDC (version for competition). As for the Faceforensics++, you should extract face using MTCNN and prepare the folder as follows:

├── faceforensics++
  └──manipulated_sequences
    ├──Deepfakes
      └──c23
        └──mtcnn
          ├──000_003
            ├──000_003_0000.png
            ├──000_003_0001.png
            ├──...
          ├──001_870
          └── ...
      └──c40
        └──mtcnn
          ├──000_003
            ├──000_003_0000.png
            ├──000_003_0001.png
            ├──...
          ├──001_870
          └── ...
    ├──Face2Face
    ├──FaceSwap
    ├──NeuralTextures

  └──original_sequences
    └──youtube
      └──c23
        └──mtcnn
          ├──000
            ├──000_0000.png
            ├──000_0001,png
            ├──...
          ├──001
          └── ...
      └──c40
        └──mtcnn
          ├──000
            ├──000_0000.png
            ├──000_0001,png
            ├──...
          ├──001
          └── ...

And change the ffpp_original_path and ffpp_fake_path in config.py as your own data_root.

Train

All the config is written in config.py. Before training, you can change the hyperparameters as you need. Then run

CUDA_VISIBLE_DEVICES = DEV_ID python train.py

to train the LTW model. The log and model are saved in folder "result" automatically.

Test

The config of the test is config_test.py. Before training, you should assign the path of tested model to the variable "model_path" in config_test.py. Then you can run

CUDA_VISIBLE_DEVICES = DEV_ID python test.py