PerdonLiu / Visual-Privacy-Protection-via-Mapping-Distortion

The implementation of our paper 'Visual Privacy Protection via Mapping Distortion', accepted by the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2021.

This is the implementation of our paper Visual Privacy Protection via Mapping Distortion, accepted by the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2021. This project is developed based on Python 3.6, created by Peidong Liu and Yiming Li.

Citation

If our work is useful for your research, please cite our paper as follows:

@inproceedings{li2021visual,
  title={Visual Privacy Protection via Mapping Distortion},
  author={Li, Yiming and Liu, Peidong and Jiang, Yong and Xia, Shu-Tao},
  booktitle={ICASSP},
  year={2021}
}

Install Pre-requisites

pip install -r requirements.txt

Generating the Modified CIFAR-10 Dataset

• Download pretrained model from Google Drive and put it in the folder "/path/to/project/runs/train_original_cifar10/checkpoints/".

• Download original cifar10 training images and original cifar10 testset from Google Drive, put them in the folder "/path/to/project/data" and unzip them.

unzip cifar10_ori_images.zip
unzip testdir.zip

• Run the script below to start the generation process for MDP.

bash generate_modified_cifar10.sh /path/to/project/experiments/generate_modified_cifar10_resnet.yaml

• Check modified cifar10 locates at '/path/to/project/data/generate_modified_cifar10_resnet/'.

• Run the script below to start the generation process for AugMDP (T=2).

bash generate_modified_cifar10.sh /path/to/project/experiments/generate_modified_cifar10_resnet_multi2.yaml

• Check modified cifar10 (T=2) locates at '/path/to/project/data/generate_modified_cifar10_resnet_multi2/'.

Training on the Modified CIFAR-10 Dataset

• Run the script below to start the training process on the dataset generated by MDP.

bash train_modified_cifar10.sh /path/to/project/experiments/train_modified_cifar10.yaml

• Run the script below to start the training process on the dataset generated by AugMDP (T=2).

bash train_modified_cifar10.sh /path/to/project/experiments/train_modified_cifar10_multi2.yaml

Acknowledgement

Thanks to MadryLab (https://github.com/MadryLab/constructed-datasets) for their inspiration.