JGBA-pointcloud-attack

Official code of MM'20 paper <Efficient Joint Gradient Based Attack Against SOR Defense for 3D Point Cloud Classification>

Requirements

torch >= 1.0
numpy
scipy
sklearn
tqdm [optional]

Dataset

We conduct experiments on a 1024-point downsampled version of ModelNet40 dataset, just as other point cloud adversarial attack papers do.

Here is the download link for the dataset:
Google Drive
Baidu Drive passwd: f9uy
Location: ./dataset/random1024/whole_data_and_whole_label.pkl

Model

Four victim classifiers are tested with, including PointNet, PointNet++ (SSG), PointNet++ (MSG), and DGCNN.

Remember to build pointnet++ before attacking it:

python setup.py build_ext --inplace

If there raise any error when you run the codes about the four models, please try to solve it by yourself before contacting us. Because we just fork the codes from their official repo 😷

Here are the download links for the checkpoints:

PointNet Google Drive, Baidu Drive passwd: ouk3
Location: ./PointNet/pointnet/cls_model_201.pth

PointNet++ (SSG) Google Drive, Baidu Drive passwd: t2zq
Location: ./PointNet2_PyTorch/checkpoints_ssg/pointnet2_cls_best.pth.tar

PointNet++ (MSG) Google Drive, Baidu Drive passwd: cdfe
Location: ./PointNet2_PyTorch/checkpoints_msg/pointnet2_cls_best.pth.tar

DGCNN Google Drive, Baidu Drive passwd: r0gc
Location: ./DGCNN/checkpoints/model.t7

How to run

untargeted attack:

python craft_adv_examples-untargeted.py --adv JGBA --eps 0.1 --n 40 --eps_iter 0.01

targeted attack:

python craft_adv_examples-targeted.py --adv JGBA --eps 0.1 --n 40 --eps_iter 0.01

Performance

The success rates of our JGBA attack on both untargeted attack and targeted attack are satisfying, because we propose to break the SOR defense directly. Please refer to the final version of our paper for more experimental results.

Citation

If you find this work useful, please consider citing our paper. We provide a BibTeX entry of our paper below:

@inproceedings{ma2020efficient,
  title={Efficient Joint Gradient Based Attack Against SOR Defense for 3D Point Cloud Classification},
  author={Ma, Chengcheng and Meng, Weiliang and Wu, Baoyuan and Xu, Shibiao and Zhang, Xiaopeng},
  booktitle={Proceedings of the 28th ACM International Conference on Multimedia},
  pages={1819--1827},
  year={2020}
}