Skip to content

xiangchong1/3d-adv-pc

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

12 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Generating 3D Adversarial Point Clouds

By Chong Xiang, Charles R. Qi, Bo Li

Sample code for CVPR 2019 paper "Generating 3D Adversarial Point Clouds" arXiv

attack pipeline

Requirement

This code is tested with Python 2.7 and Tensorflow 1.10.0

Other required packages include numpy, joblib, sklearn, etc.

Usage

There are four Python scripts in the root directorty for different attacks:

  • perturbation.py -- Adversarial Point Pertubations
  • independent.py -- Adversarial Independent Points
  • cluster.py -- Adversarial Clusters
  • object.py -- Adversarial Objects

The code logics of these four scripts are similar; they attack the victim objects into the specified target class. The basic usage is python perturbation.py --target=5.

Other parameters can be founded in the script, or run python perturbation.py -h. The default parameters are the ones used in the paper.

Other files

  • log/model.ckpt -- the victim model used in the paper. Download link.
  • data/attacked_data.z -- the victim data used in the paper. It can be loaded with joblib.load, resulting in a Python list whose element is a numpy array (shape: 25*1024*3; 25 objects of the same class, each object is represented by 1024 points)
  • gen_initial.py -- used to generate initial points for adversarial cluster/object. The script uses DBSCAN to cluster the generated critical points.
  • critical -- the default directory to dump the generated initial points
  • data/airplane.py -- the airplane object used in the paper as a uav for the adversarial object. can be loaded with np.load.
  • utils/tf_nndistance -- a self-defined tensorlfow op used for Chamfer/Hausdorff distance calculation. Use tf_nndistance_compile.sh to compile the op. The bash code may need modification according to the version and installtion path of CUDA. Note that it should be OK to directly calculate Chamfer/Hausdorff distance with available tf ops instead of tf_nndistance.

Misc

  • The sample adversarial point clouds can be downloaded here. The targeted model is log/model.ckpt
  • The aligned version of ModelNet40 data (in point cloud data format) can be downloaded here.
  • The visulization in the paper is rendered with MeshLab
  • Please open an issue or contact Chong Xiang ([email protected]) if there is any question.

About

Generating 3D Adversarial Point Clouds

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published