lvpeizhuo/DBIA

This is the source code for DBIA. Our paper is accepted by IEEE ICME2023.

The code for DBIA on DeiT

• This is the source code of DBIA: Data-free Backdoor Injection Attack against Transformer Networks https://arxiv.org/pdf/2111.11870.pdf

• Moreover, we refer to the released code of DeiT, which is from https://github.com/facebookresearch/deit

Environment

• PyTorch 1.7.0+ and torchvision 0.8.1+ and pytorch-image-models 0.3.2

Data Preparation

• Imagenet2012 validation set data is placed in ./data/imagenet2012/val

Starting Poisoning

• Generate the background picture required for training trigger

    python preprocess.py

• Background picture and blend_ tensor.pt is stored in ./savedfigure

Reversing Trigger

• First, write the path of blend_tensor.pt on line 38 of the file activate_trigger.py

    python activate_trigger.py

• Trigger pictures are stored in the triggers folder

Training Poisoning Model

• First, write the path of the selected trigger on line 35 of the file tensors_dataset.py

    python backdoor_main.py

Citation

• If it helps you, you can cite our paper as below:

    @article{lv2021dbia,
    title={Dbia: Data-free backdoor injection attack against transformer networks},
    author={Lv, Peizhuo and Ma, Hualong and Zhou, Jiachen and Liang, Ruigang and Chen, Kai and Zhang, Shengzhi and Yang, Yunfei},
    journal={arXiv preprint arXiv:2111.11870},
    year={2021}
  }