Geometry-Aware Generation of Adversarial Point Clouds

By Yuxin Wen, Jiehong Lin, Ke Chen, C. L. Philip Chen, Kui Jia.

Introduction

This repository contains the implementation of our paper https://arxiv.org/abs/1912.11171. (This documentation is still under construction, please refer to our paper for more details)

Requirements

• A computer running on Linux
• NVIDIA GPU and NCCL
• Python 3.6 or higher version
• Pytorch 1.1 or higher version

Usage

Model tranining and data preparing

Use python main.py to train a new model. Here is an example settings for PointNet:

python main_train.py --datadir /data/modelnet40_normal_resampled/ --npoint 1024 --arch PointNet --epochs 200

Note that /data/modelnet40_normal_resampled/ is the path of your ModelNet40 dataset. We use the dataset (ModelNet40) of PointNet++ which can be download here.

Before running the attack, you can scale down the data and involves only the instances you want to attack, here is an example:

python Provider/gen_data_mat.py --datadir /data/modelnet40_normal_resampled/ --npoint 1024 --arch PointNet --out_datadir Data/ --out_classes 10 --max_out_num 25

And then the .mat file with 250 instances from 10 different classes would be generated, of which all are correctly classified.

If you DO NOT want to generate the .mat file yourself, you can download one here, for the pretrained network provided in Pretrained/PointNet/1024/.

Attack

Use python attack.py to generate adversarial point clouds:

python main_attack.py --data_dir_file Data/modelnet10_250instances1024_PointNet.mat --npoint 1024 --arch PointNet \
--attack GeoA3 --attack_label All --binary_max_steps 10 --iter_max_steps 500 \
--cls_loss_type CE --dis_loss_type CD --dis_loss_weight 1.0 --hd_loss_weight 0.1 --curv_loss_weight 1.0 --curv_loss_knn 16 \
--lr 0.01

Defense

defense.py is used for evaluating the defense results on the corresponding adversarial point clouds:

python defense.py --datadir Exps/PointNet_npoint1024/All/Pertub_0_BiStep10_IterStep500_Optadam_Lr0.01_Initcons10_CE_CDLoss1.0_HDLoss0.1_CurLoss1.0_k16/Mat \
	--npoint 1024 --arch PointNet \
	--defense_type outliers_fixNum --drop_num 128

Citation

If you use this method or this code in your paper, then please cite it:

@ARTICLE{9294112,
  author={Y. {Wen} and J. {Lin} and K. {Chen} and C. L. P. {Chen} and K. {Jia}},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
  title={Geometry-Aware Generation of Adversarial Point Clouds}, 
  year={2020},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/TPAMI.2020.3044712}
}