CLEANIR: Controllable Attribute-Preserving Natural Identity Remover

This is the official implementation of CLEANIR: Controllable Attribute-Preserving Natural Identity Remover

Requirements

• face-recognition>=1.2.3
• opencv-python>=4.1.0.25
• Keras>=2.2.4
• tensorflow-gpu>=1.13.0rc0+nv
• matplotlib>=3.0.2
• numpy>=1.14.5
• pandas>=0.23.0
• tqdm>=4.32.2
• gdown>=3.10.2
• keras-facenet>=0.1a5 (if you want to run codes for evaluation on de-identification)
• azure-cognitiveservices-vision-face>=0.4.0 (if you want to run codes for evaluation on preserving facial emotion)

Testing environment

• Ubuntu 16.04
• Python 3.5.2
• Keras 2.2.4 (backend=TensorFlow 1.13.0-rc0)

Usage

Please check CLEANIR_notebook.ipynb for testing and CLEANIR_train_notebook.ipynb for training.

Citation

If you find this work useful for your research, please consider citing our paper:

@article{cho2020cleanir,
  title={CLEANIR: Controllable Attribute-Preserving Natural Identity Remover},
  author={Cho, Durkhyun and Lee, Jin Han and Suh, Il Hong},
  journal={Applied Sciences},
  volume={10},
  number={3},
  pages={1120},
  year={2020},
  publisher={Multidisciplinary Digital Publishing Institute}
}