KO | EN

GitLite

Minimalist Open Source Repository Search Engine

gitlite — search

#typescript #ai-agents #developer-tools #cli #codex #claude-code #react #dsh-plugin #mcp #open-source #dsh #deepseek-harness

GeoA3

★ 35 Open GitHub ↗

Code for Geometry-Aware Generation of Adversarial Point Clouds

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# 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++](https://github.com/charlesq34/pointnet2) which can be download [here](https://shapenet.cs.stanford.edu/media/modelnet40_normal_resampled.zip).

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](https://drive.google.com/file/d/1mFsEyvfetQDlA30pHijN3S1wAlhwuemk/view?usp=sharing), 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}
}
```

GeoA3

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README