Continual Face Forgery Detection via Historical Distribution Preserving.

This repository is the official implementation of "Continual Face Forgery Detection via Historical Distribution Preserving."

Dependencies and Installation

Environment

• Python 3 (Recommend to use Anaconda)
• PyTorch >= 1.0
• NVIDIA GPU + CUDA (Need at least 15G GPU memory)

Before run the code, you should run

$ pip install -r requirements.txt

Dataset

This code is mainly focus on : Faceforensics++ (four subsets contains Deepfakes, Face2Face, FaceSwap, NeuralTextures), CELEB-DF, DFDC and WildDeepfake. And change the data_root in configs/test.yaml and configs/train_hdp.yaml for each dataset. Here we give the example for Protocol 1. For other Protocols or other continue setting, you can add your own dataloader.py and change the train_hdp.yaml.

As for the Faceforensics++, you should extract face using face detector and prepare the folder as follows:

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

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

Train

The training and testing config is written in configs/train_hdp.yaml and configs/test.yaml. Before training, you can change the hyperparameters as you need. Then run

CUDA_VISIBLE_DEVICES=x python3 train.py -c configs/train_hdp.yaml

to train the model. The id, model, log, generated UAP are saved in folder "wandb" automatically.

You can run

CUDA_VISIBLE_DEVICES=x python3 train_baseline.py -c configs/train_hdp.yaml

to get sequence finetune result as baseline.

Test

After training, you can run

CUDA_VISIBLE_DEVICES = DEV_ID python3 test.py --exam_id XXX -c configs/test.yaml

to evaluate the final stage model with specific id on all testsets.

Cite

If you are interested in this work, please cite our paper

@article{sun2024continual,
  title={Continual Face Forgery Detection via Historical Distribution Preserving},
  author={Sun, Ke and Chen, Shen and Yao, Taiping and Sun, Xiaoshuai and Ding, Shouhong and Ji, Rongrong},
  journal={International Journal of Computer Vision},
  pages={1--18},
  year={2024},
  publisher={Springer}
}