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++.

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. 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, and its code is available here.

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.

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