[CVPR2020] On Isometry Robustness of Deep 3D Point Cloud Models under Adversarial Attacks

Environment

Ubuntu 16.04.5 LTS
GPU RTX2080ti
Python 3.7
Install the python dependencies with

pip install -r requirements.txt

Data

• [ModelNet40] automatically downloaded
• [ShapeNet] /fxia22/pointnet.pytorch (follow the guidence for downloading)
  The default path of data is '/data'.

Usage Sample

Train model

With default parameters setting, run

python train.py --data modelnet40 --model pointnet

Trained model is stored in '/checkpoints' with log in '/logs_train'.

Launch attack

If you don't want to retrain the model, download a trained model here (with ModelNet40 data, PointNet model), move it to '/checkpoints', then run

python attack.py --data modelnet40 --model pointnet --model_path 'example'

The attack log is stored in '/logs_attack'. The attack is default to be CTRI since TSI is done at the same time.

Supplementary materials

Please check here for supplementary materials mentioned in this paper.

References

• PointNet /charlesq34/pointnet, /fxia22/pointnet.pytorch
• PointNet++ /charlesq34/pointnet2, /yanx27/Pointnet_Pointnet2_pytorch
• DG-CNN /WangYueFt/dgcnn
• RS-CNN /Yochengliu/Relation-Shape-CNN
• Thompson Sampling /andrecianflone/thompson
• Adversarial Attacks /MadryLab, /YyzHarry/ME-Net