This is the official repository for "How to Craft Backdoors with Unlabeled Data Alone?" by Yifei Wang*, Wenhan Ma*, Stefanie Jegelka, Yisen Wang.

torch==1.10.0
torchvision==0.11.1
einops
pytorch-lightning==1.5.3
torchmetrics==0.6.0
lightning-bolts>=0.4.0
tqdm
wandb
scipy
timm

Split training set into pretraining set and downstream set. Run python ./misc/cifar10_split.py

Pretrain the clean encoder on pretraining set. Run the script in script/cifar10_encoder

Select the poison subset. Run the script in script/cifar10_poison

Pretrain the backdoor encoder on pretraining set with poison subset. Run the script in script/cifar10_pretrain

Train the classifier. Run the script in script/cifar10_linear

Split training set into pretraining set and downstream set. Run

python ./misc/imagnet_script/in100_split.py

Create a copy with poison. Run

python ./misc/imagnet_script/in100_add_trigger.py

Create a folder with soft link to the clean dataset. Run

python misc/imagnet_script/in100_link.py

Pretrain the clean encoder on pretraining set. Run the script in script/cifar10_encoder

Select the poison subset. Run the script in script/cifar10_poison

Create a folder with soft link to the clean dataset and poison subset. Run

python misc/imagnet_script/in100_link_poison.py

Pretrain the backdoor encoder on pretraining set with poison subset. Run the script in script/cifar10_pretrain

Train the classifier. Run the script in script/cifar10_linear

If you find this useful in your research, please consider citing:

@misc{wang2024craft,
      title={How to Craft Backdoors with Unlabeled Data Alone?}, 
      author={Yifei Wang and Wenhan Ma and Stefanie Jegelka and Yisen Wang},
      year={2024},
      eprint={2404.06694},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

This repo is based upon the following repository:

• solo-learn