Awesome

Shape-invariant 3D Adversarial Point Clouds (CVPR 2022)

This repository provides the official PyTorch implementation of the following conference paper:

Shape-invariant 3D Adversarial Point Clouds (CVPR 2022) <br> Qidong Huang¹, Xiaoyi Dong¹, Dongdong Chen², Hang Zhou³, Weiming Zhang¹, Nenghai Yu¹ <br> ¹University of Science and Technology of China, ²Microsoft Cloud AI, ³Simon Fraser University <br>

TL;DR quick start

To setup a conda environment, download the required dataset, and begin the attack process:

conda env create -f environment.yml
conda activate si_adv_pc
bash download.sh
python main.py --transfer_attack_method ifgm_ours 

Setup

The code is tested with Python3.7, Pytorch >= 1.6 and CUDA >= 10.2, including the following dependencies:

• tqdm >= 4.52.0
• numpy >= 1.19.2
• scipy >= 1.6.3
• open3d >= 0.13.0
• torchvision >= 0.7.0
• scikit-learn >= 1.0

When running the code on CurveNet for the first time, it will take a few minutes to automatically compile the cuda extension. Here we list our dependencies for reference:

• gcc == 9.4.0
• ninja == 1.7.2

Preparation

Download the aligned ModelNet40 dataset and ShapeNetPart dataset in their point cloud format and unzip them into your own dataset path.

Download the pretrained models we provided for attack evaluation and unzip them at ./checkpoint. The available models include PointNet, PointNet++, DGCNN, PointCNN, PointConv, RS-CNN, PCT, PAConv, SimpleView and CurveNet.

White-box attack

To implement white-box shape-invariant attack on PointNet, please run the following command:

python main.py 
--dataset ModelNet40
--data_path /your/path/to/dataset/
--transfer_attack_method ifgm_ours 
--surrogate_model pointnet_cls 
--target_model pointnet_cls 
--step_size 0.007 
--max_steps 50 
--eps 0.16

Black-box transfer-based attack

To implement transfer-based attack from PointNet to black-box target CurveNet, please run the following command:

python main.py 
--dataset ModelNet40
--data_path /your/path/to/dataset/
--transfer_attack_method ifgm_ours 
--surrogate_model pointnet_cls 
--target_model curvenet
--step_size 0.007 
--max_steps 50 
--eps 0.16

Black-box query-based attack

To implement black-box query-based shape-invariant attack, if you choose DGCNN as the surrogate model to attack target black-box PAConv, please run the following command:

python main.py 
--dataset ModelNet40
--data_path /your/path/to/dataset/
--query_attack_method ours 
--surrogate_model dgcnn 
--target_model paconv
--step_size 0.32

Citation

If you find this work useful for your research, please cite our paper:

@inproceedings{huang2022siadv,
  title={Shape-invariant 3D Adversarial Point Clouds},
  author={Qidong Huang and Xiaoyi Dong and Dongdong Chen and Weiming Zhang and Nenghai Yu},
  journal={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2022}
}

License

The code is released under MIT License (see LICENSE file for details).