https://github.com/Daniel-Liu-c0deb0t/3D-Neural-Network-Adversarial-Attacks

Research on adversarial attacks and defenses for deep neural network 3D point cloud classifiers like PointNet and PointNet++.
https://github.com/Daniel-Liu-c0deb0t/3D-Neural-Network-Adversarial-Attacks

3d-point-clouds adversarial-attacks algorithms deep-learning matplotlib numpy pointnet python tensorflow

Last synced: 3 months ago
JSON representation

Research on adversarial attacks and defenses for deep neural network 3D point cloud classifiers like PointNet and PointNet++.

• Host: GitHub
• URL: https://github.com/Daniel-Liu-c0deb0t/3D-Neural-Network-Adversarial-Attacks
• Owner: Daniel-Liu-c0deb0t
• Created: 2018-07-15T19:15:52.000Z (over 6 years ago)
• Default Branch: master
• Last Pushed: 2020-05-22T05:40:50.000Z (over 4 years ago)
• Last Synced: 2024-05-29T05:13:26.422Z (6 months ago)
• Topics: 3d-point-clouds, adversarial-attacks, algorithms, deep-learning, matplotlib, numpy, pointnet, python, tensorflow
• Language: Python
• Homepage:
• Size: 82 KB
• Stars: 23
• Watchers: 5
• Forks: 9
• Open Issues: 4
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

• awesome-3D-point-cloud-attacks - [Code

README

        
# 3D Point Cloud Adversarial Attacks and Defenses

Adversarial attacks and defenses on neural networks that process 3D point cloud data, namely PointNet and PointNet++. The preprint paper is available on Arxiv [**here**](https://arxiv.org/abs/1901.03006). A shortened version is accepted at the 2019 IEEE ICIP. If you use this code, please cite

```

@article{liu2019extending,

  title={Extending Adversarial Attacks and Defenses to Deep 3D Point Cloud Classifiers},

  author={Liu, Daniel and Yu, Ronald and Su, Hao},

  journal={arXiv preprint arXiv:1901.03006},

  year={2019}

}

```

A continuation of this research is in [this paper](https://arxiv.org/abs/1908.06062), and its code is available [here](https://github.com/Daniel-Liu-c0deb0t/Adversarial-point-perturbations-on-3D-objects).

Note that files modified from the PointNet and PointNet++ source codes are included. Some files may need to be moved to the correct location before running experiments. Clone the PointNet and PointNet++ repositories, train the models, and place the code in `src/` into each of those repositories. The code in the `src/pointnet` and `src/pointnet2` directories have to replace the corresponding files in the PointNet and PointNet++ repositories. The files `src/adversarial_pointnet.py` and `src/adversarial_pointnet2.py` should be ran for attacks and defenses. Get the `.npz` file required for gradient projection attacks from [here](https://github.com/Daniel-Liu-c0deb0t/Adversarial-point-perturbations-on-3D-objects/releases/download/Data/point_clouds.npz).

## Highlights

### Attacks

- Fast/iterative gradient sign

- Jacobian-based saliency map attack

- Gradient projection

- Clipping L2 norms

### Defenses

- Adversarial training

- Outlier removal

- Salient point removal

### Conclusions

- Adversarial attacks are effective against deep 3D point cloud classifiers

- It is more easy to defend point cloud classifiers than 2D image classifiers