AdvCloak

Code and datasets of paper AdvCloak: Customized Adversarial Cloak for Privacy Protection.

Usage Instructions

Environment

Please install Anaconda, Pytorch and MxNet. For other libs, please refer to the file requirements.txt.

conda create -n AdvCloak python=3.8
conda activate AdvCloak
git clone https://github.com/liuxuannan/AdvCloak.git
pip install -r requirements.txt

Datasets and face recognition models

• You can download the Privacy-Commons dataset and Privacy-Celebrities dataset using Baidu Netdisk(password: 3g2b), Google Drive; and Privacy-Celebrities dataset Baidu Netdisk(password: 28cq), Google Drive.

• Create a folder ['code/data/'], and then unzip the datasets into it.

• Please download Source models and Target models: Baidu Netdisk(password: y1cy), Google Drive.

• Create a folder ['code/generation/source_mdoel/'], and then unzip the Source models into it. Create a folder ['code/evaluation/target_mdoel/'], and then unzip the target models into it.

Pre-trained Stage 1 and Stage 2 checkpoint for AdvCloak

You can download the pre-trained checkpoing of AdvCloak model at the first stage training using this link

• Create a folder ['code/generation/stage_1'], and then unzip the checkpoint of stage 1 into it.

You can download the final checkpoint of AdvCloak model with two-stage training using this link

• Create a folder ['code/generation/stage_2'], and then unzip the checkpoint of final AdvCloak into it.

Privacy Mask Generation

To generate privacy masks of Privacy-Commons dataset, based on surrgate model "Resnet50-WebFace-ArcFace", with different approximation methods, and transferability enhancement methods, please do as follows. Other surrogate models can be used modifying "--pretrained". Other parameters, please refer to the code.

cd code/generation
python mask_generation.py --query_image_dir ../data  --query_train_image_list  ../data/list/privacy_train_v3_10.lst  --pretrained_generator  ./models/stage_2_model/AdvCloak.pth  --mask_out  ./mask_out

To generate privacy masks of Privacy-Celebrities dataset, please do as follows.

cd code/generation
python mask_generation.py --query_image_dir ../data  --query_train_image_list  ../data/list/privacy_ms90w_train.lst  --pretrained_generator  ./models/stage_2_model/AdvCloak.pth  --mask_out  ./mask_out

Privacy Mask Evaluation

After generating the privacy masks, please refer to the evaluation part for privacy pretection rate. You can modify "--msk_dir" for different versions of masks. For Privacy-Commons dataset, evaluation towards six target models is as follows.

cd code/evaluation
./test_common.sh 

For Privacy-Celebrities dataset, evaluation towards six target models is as follows.

./test_celeb.sh

Citation

If you find AdvCloak useful in your research, please consider to cite:

@article{liu2023advcloak,
    title={Advcloak: Customized adversarial cloak for privacy protection},
    author={Liu, Xuannan and Zhong, Yaoyao and Cui, Xing and Zhang, Yuhang and Li, Peipei and Deng, Weihong},
    journal={arXiv preprint arXiv:2312.14407},
    year={2023}
}