Awesome

This is the implementation of our paper 'Open-sourced Dataset Protection via Backdoor Watermarking', accepted by the NeurIPS Workshop on Dataset Curation and Security, 2020. [arXiv] This project is developed based on Python 3.6, created by Yiming Li and Ziqi Zhang.

Citation

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

@inproceedings{li2020open,
  title={Open-sourced Dataset Protection via Backdoor Watermarking},
  author={Li, Yiming and Zhang, Ziqi and Bai, Jiawang and Wu, Baoyuan and Jiang, Yong and Xia, Shu-Tao},
  booktitle={NeurIPS Workshop},
  year={2020}
}

Install prerequisites

pip install -r requirements.txt

Experiments on CIFAR-10 dataset

All code for this part are included in the "CIFAR" subfolder. Please change to that folder before running the code.

Standard Training

ResNets(-18)

python train_standard.py --gpu-id 0 --checkpoint 'checkpoint/benign'

VGG(-19)

python train_standard_vgg.py --gpu-id 0 --checkpoint 'checkpoint/benign_vgg'

Training with Watermarked Dataset

ResNets(-18)

The trigger is a white 3*3 square on the bottom right corner of the image with transparency = 1 (poisoning rate is set to 0.1).

python train_watermarked.py --gpu-id 0 --poison-rate 0.1 --checkpoint 'checkpoint/infected/square_1_01' --trigger './Trigger_default1.png' --alpha './Alpha_default1.png'

The trigger is a 3-pixel-width black horizontal line above the image with transparency = 1 (poisoning rate is set to 0.1).

python train_watermarked.py --gpu-id 0 --poison-rate 0.1 --checkpoint 'checkpoint/infected/line_1_01' --trigger './Trigger_default2.png' --alpha './Alpha_default2.png'

VGG(-19)

The trigger is a white 3*3 square on the bottom right corner of the image with transparency = 1 (poisoning rate is set to 0.1).

python train_watermarked_vgg.py --gpu-id 0 --poison-rate 0.1 --checkpoint 'checkpoint/infected_vgg/square_1_01' --trigger './Trigger_default1.png' --alpha './Alpha_default1.png'

The trigger is a 3-pixel-width black horizontal line above the image with transparency = 1 (poisoning rate is set to 0.1).

python train_watermarked_vgg.py --gpu-id 0 --poison-rate 0.1 --checkpoint 'checkpoint/infected_vgg/line_1_01' --trigger './Trigger_default2.png' --alpha './Alpha_default2.png'

Dataset Verification with Pairwise T-test

The following is an example of the verification with trigger1, alpha1, margin=0.2 under ResNets structure.

python test_cifar.py --gpu-id 0 --model 'resnet' --trigger './Trigger_default1.png' --alpha './Alpha_default1.png' --margin 0.2 --model-path './checkpoint/infected/line_1_01/checkpoint.pth.tar'

Experiments on GTSRB dataset

All code for this part are included in the "GTSRB" subfolder. Please change to that folder and download the dataset [link] before running the code.