DefakeHop: A Light-Weight High-Performance Deepfake Detector

This is the official Python implementation of our work: "DefakeHop: A Light-Weight High-Performance Deepfake Detector" accepted at ICME 2021.

State-of-the-art Deepfake detection methods are built upon deep neural networks. In this work, we proposed a non deep learning method to detect Deepfake videos which use the successive subspace learning (SSL) principle to extract features from various parts of face images. The features are also further distilled by our feature distillation module to derive a concise representation of the fake and real faces.

Required packages

conda install -c anaconda pandas 
conda install -c conda-forge opencv
conda install -c conda-forge xgboost
conda install -c anaconda scikit-image
conda install -c conda-forge matplotlib
conda install -c conda-forge scikit-learn

Preprocessing

• Extracting the facial landmarks using OpenFace. Please check here more more details.

FeatureExtraction -f [video path] -out_dir [output directory]

• Face alignment and Crop the facial regions

python preprocessing/extract_facial_regions.py [videos folder] [landmarks folder] [output directory]

How to run

We use UADFV dataset as an example to show how to use our code to train and test the model.

python model.py

When we train the model, we use three items to train.

• Images: 4D numpy array (N,H,W,C).

• Labels: 1D numpy array where 1 is Fake and 0 is Real.

• Names: 1D numpy array storing frame names.

  The frame name should follow the format of {video_name}_{frame_number}.

  Example: real/0047_0786.bmp, we can know it is the 786 th frame from real/0047.mp4

Cite us

If you use this repository, please consider to cite.

@misc{chen2021defakehop,
      title={DefakeHop: A Light-Weight High-Performance Deepfake Detector}, 
      author={Hong-Shuo Chen and Mozhdeh Rouhsedaghat and Hamza Ghani and Shuowen Hu and Suya You and C. -C. Jay Kuo},
      year={2021},
      eprint={2103.06929},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Acknowledgment

This work was supported by the Army Research Laboratory (ARL) under agreement W911NF2020157.