AdvGCGAN

The code repository for our paper Generating Unrestricted 3D Adversarial Point Clouds arXiv

Installation

This repository is based on Python 3.8, Pytorch 1.8.1, CUDA 11.2 on Ubuntu 18.04.

1. Set up environments for the codes.

   pip install -r requirements.txt

2. (Optional) Install KNN_CUDA.

   pip install --upgrade https://github.com/unlimblue/KNN_CUDA/releases/download/0.2/KNN_CUDA-0.2-py3-none-any.whl

3. (Optional) Compile Structural_Losses operators by using the Makefile located at evaluation/pytorch_structural_losses

4. Download ShapeNetCore data set and unzip.

   wget -c https://shapenet.cs.stanford.edu/media/shapenetcore_partanno_segmentation_benchmark_v0.zip
   unzip shapenetcore_partanno_segmentation_benchmark_v0.zip

5. Download the pretrained PointNet, PointNet++, and DGCNN model from GoogleDrive, copy them to the folder checkpoints/shapenetpart_pointnet_train/, checkpoints/shapenetpart_pointnet2_train/, and checkpoints/shapenetpart_dgcnn_train/. We also give some of our pre-trained files.

Usage

1. Parameter Settings:

   You should adjust the parameters for the training of AdvGCGAN, they can be found in arguments.py.

   You should specify the classes for the traning. In our paper, we use ['Airplane', 'Table','Car'] for the class_choice .

   We suggest only train with 2 classes due to the limited performance of the Generator.

   After the GAN training, you should adjust the class_choice to ['Table'] if you want to generate adversarial point clouds for 'Table' class.

2. Run:

   According to the paper, there are two stage training. Please use train.py for GAN training, and use train-attack.py for adversarial training.

   python train.py
   python train-attack.py

3. Evaluate

   We also provide the evaluation file to test the performance of different attack models. You should follow the two (Optioanl) steps for using our provided codes. The evaluation contains SRS and SOR defense methods. You also should adjust the parameters in arguments.py.

   python evaluate.py

Acknowledgement

The basic structure of our codes is adopted from the paper 3D Point Cloud Generative Adversarial Network Based on Tree Structured Graph Convolutions (Dong Wook Shu*, Sung Woo Park*, Junseok Kwon)

Reference

Please cite our paper if you found any helpful information:

@article{dai2021generating,
  title={Generating Unrestricted 3D Adversarial Point Clouds},
  author={Dai, Xuelong and Li, Yanjie and Dai, Hua and Xiao, Bin},
  journal={arXiv preprint arXiv:2111.08973},
  year={2021}
}