• Table of Contents
  • Overview
  • Code Architecutre
  • Environments
  • Experiments
    • Usage
    • Configurations
    • Outputs
  • Citation
  • Acknowledgement

• This is the PyTorch implementation for CVPR 2024 paper "LOTUS: Evasive and Resilient Backdoor Attacks through Sub-Partitioning".
• [arXiv] | [video] | [slides] | [poster]

.
├── checkpoint        # Saved models
├── data              # Dataset folder
├── models            # Model architectures
│   ├── resnet.py     # ResNet models
│   └── vgg.py        # VGG models
├── dataset.py        # Dataset processing function
├── main.py           # Main function
├── partition.py      # (Implicit) partioning function
├── train.py          # Training function
├── trigger.py        # Trigger function
└── utils.py          # Utility functions

# Create python environment (optional)
conda env create -f environment.yml
source activate lotus

We provide example code snippets for CIFAR-10 dataset. These can be easily plugged in and modified in ./utils.py, specifically within the get_dataset(*) functions.

To train and evaluate a backdoored model using LOTUS, run:

python main.py --gpu 0

It involves three steps to launch LOTUS:

• Step 1 (Line 32): Train a clean model.
• Step 2 (Line 35): Train a surrogate model for partitioning.
• Step 3 (Line 38): Poison the model with backdoor triggers.

The specific arguments and hyperparameters used to launch LOTUS can be found in ./main.py, particularly in lines 45-61.

Several output files will be saved in the directory ./checkpoint.

• ./checkpoint/clean.pt: Clean model.
• ./checkpoint/lotus_best.pt: Best backdoored model.
• ./checkpoint/lotus_final.pt: Final backdoored model.
• ./checkpoint/result.json: ASRs of different combinations of trigger and partition.
• ./checkpoint/surrogate.pt: Surrogate model for partitioning.
• ./checkpoint/training.log: Training logs.

Please cite our paper if you find it useful for your research.😀

@inproceedings{cheng2024lotus,
  title={Lotus: Evasive and resilient backdoor attacks through sub-partitioning},
  author={Cheng, Siyuan and Tao, Guanhong and Liu, Yingqi and Shen, Guangyu and An, Shengwei and Feng, Shiwei and Xu, Xiangzhe and Zhang, Kaiyuan and Ma, Shiqing and Zhang, Xiangyu},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={24798--24809},
  year={2024}
}

• BackdoorVault
• OrthogLinearBackdoor
• FLIP