Invisible Backdoor Attack with Sample-Specific Triggers

Environment

This project is developed with Python 3.6 on Ubuntu 18.04. Please run the following script to install the required packages

pip install -r requirements.txt

Demo

Before running the code, please download the checkpoints from Baidudisk (code:o89z), and put them into ckpt folder.

1. Generating poisoned sample with sample-specific trigger.

   # TensorFlow
   python encode_image.py \
   --model_path=ckpt/encoder_imagenet \
   --image_path=data/imagenet/org/n01770393_12386.JPEG \
   --out_dir=data/imagenet/bd/ 
   

   Benign image	Backdoor image	Trigger

2. Runing test.py for testing benign and poisoned images.

   # PyTorch
   python test.py
   

Train

1. Download data from Baidudisk(code:oxgb) and unzip it to folder datasets/.

2. Run training script bash train.sh.

3. The files in checkpoint folder are as following:

   --- args.json # Input arguments
   |-- x_checkpoint.pth.tar # checkpoint
   |-- x_model_best.pth.tar # best checkpoint
   |-- x.txt # log file
   

Defense

Check BackdoorBench for details

Citation

Please cite our paper in your publications if it helps your research:

@inproceedings{li_ISSBA_2021,
  title={Invisible Backdoor Attack with Sample-Specific Triggers},
  author={Li, Yuezun and Li, Yiming and Wu, Baoyuan and Li, Longkang and He, Ran and Lyu, Siwei},
  booktitle={IEEE International Conference on Computer Vision (ICCV)},
  year={2021}
}

Notice

This repository is NOT for commecial use. It is provided "as it is" and we are not responsible for any subsequence of using this code.