PointCRT

The implementation of our ACM MM 2023 paper "PointCRT: Detecting Backdoor in 3D Point Cloud via Corruption Robustness "[paper]

Requirements

• torch==1.13.0
• xgboost
• scikit-learn
• open3d

Getting Started

• Clone this Code Repository

git clone https://github.com/CGCL-codes/PointCRT.git
cd PointCRT

• Build Third-party modules Pygem

Some corruption methods in ModelNet40-C require it. More installation details could be found here.

git clone https://github.com/mathLab/PyGeM.git
python setup.py install

• Update the file path of the datasets in config.yaml

ModelNet_path: 		'/Path/to/ModelNet/'
ShapeNetPart_path: 	'/Path/to/shapenet/'

Quick start

Generate backdoor sample

Similar approaches can be considered for other backdoor attack methods. However, it should be noted that `SC`` is optional. This is due to the requirement in the PCBA algorithm.

python PCBA.py --target_model pointnet_cls --dataset ModleNet40 --SC 8 --TC 35

• attack_dir: /path/to/backdoor/sample
• SC: index of source class
• TC index of target class

Train a Backdoor Model

python train_attacked.py --target_model pointnet_cls--dataset ModleNet40 --attack_dir attack --output_dir model_attacked

• attack_dir: /path/to/backdoor/sample
• output_dir: /path/to/results

Detection the Backdoor sample

sh script/Process.sh
python PointCRT.py --result_dir model_attacked --target_model pointnet_cls --dataset ModelNet40

You can also customize above procedure by modifying the configuration of Process.sh

Evaluate the clean models

If you want to evaluate the accuracy of the clean models, you can utilize the following pre-trained models for evaluation. The Google Drive link is here.

The pre-trained models are provided by SI-Adv. The available models include PointNet, PointNet++, DGCNN, PointCNN, PointConv, RS-CNN, PCT, SimpleView and CurveNet.

BibTex

@inproceedings{hu2023pointcrt,
  author = {Hu, Shengshan and Liu, Wei and Li, Minghui and Zhang, Yechao and Liu, Xiaogeng and Wang, Xianlong and Zhang, Leo Yu and Hou, Junhui},
  title = {PointCRT: Detecting Backdoor in 3D Point Cloud via Corruption Robustness},
  doi = {10.1145/3581783.3612456},
  booktitle = {Proceedings of the 31st ACM International Conference on Multimedia},
  year = {2023}
}

Reference

This codebase is based on PCBA, and the corruption methods are based on ModelNet40-C.