Skip to content

ajhamdi/AdvPC

Repository files navigation

AdvPC: Transferable Adversarial Perturbations on 3D Point Clouds (ECCV 2020)

By Abdullah Hamdi, Sara Rojas , Ali Thabet, Bernard Ghanem

The official code of ECCV 2020 paper "AdvPC: Transferable Adversarial Perturbations on 3D Point Clouds". We perform transferable adversarial attacks on 3D point clouds by utilizing a point cloud autoencoder. we exceed SOTA by up to 40% on transferability and 38% in breaking SOTA 3D defenses on ModelNet40 data.

attack pipeline

Citation

If you find our work useful in your research, please consider citing:

@InProceedings{10.1007/978-3-030-58610-2_15,
author="Hamdi, Abdullah
and Rojas, Sara
and Thabet, Ali
and Ghanem, Bernard",
editor="Vedaldi, Andrea
and Bischof, Horst
and Brox, Thomas
and Frahm, Jan-Michael",
title="AdvPC: Transferable Adversarial Perturbations on 3D Point Clouds",
booktitle="Computer Vision -- ECCV 2020",
year="2020",
publisher="Springer International Publishing",
address="Cham",
pages="241--257",
isbn="978-3-030-58610-2"
}

Requirement

This code is tested with Python 2.7 and Tensorflow 1.9/1.10

Other required packages include numpy, joblib, sklearn, etc.( see environment.yml)

creating conda environment and compiling tf_ops C++ libraries

  • conda create -n NAME python=2.7 anaconda
  • conda activate NAME
  • conda install tensorflow-gpu=1.10.0
  • conda install -c anaconda cudatoolkit==9
  • make sure CUDA/Cudnn is there by running nvcc --version , gcc --version, whereis nvcc
  • look for TensorFlow paths in your device, it should be something like this /home/USERNAME/.local/lib/python2.7/site-packages/tensorflow
  • change TF_INC,TF_LIB,nsync in the makefile file in latent_3d_points/external/structural_losses/ according to the above TF path
  • run make inside the above the directory

Usage

There are two main Python scripts in the root directorty:

  • attack.py -- AdvPC Adversarial Point Pertubations
  • evaluate.py -- code to evaluate the atcked point clouds under different networks and defeneses

To run AdvPC to attack network NETWORK and also evaluate the the attack, please use the following command:

python attack.py --phase all --network NETWORK --step=1 --batch_size=5 --num_iter=100 --lr_attack=0.01 --gamma=0.25 --b_infty=0.1 --u_infty=0.1 --evaluation_mode=1
  • NETWORK is one of four networks : PN: PointNet, PN1:PointNet++ (MSG) , PN2: PointNet++ (SSG), GCN: DGCNN
  • b_infty , u_infty is the L_infty norm budget used in the experiments.
  • step is the number of different initilizations for the attack.
  • lr_attack is the learning rate of the attack.
  • gamma is the main hyper parameter of AdvPC (that trades-off success with transferablity).
  • num_iter is the number of iterations in the optimzation.
  • evaluation_mode is the evaluation mode of the attack (0:targeted , 1:untargeted)

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

The results will be saved in results/exp0/ with the original point cloud and attacked point cloud saved as V_T_B_orig.npy and V_T_B_adv.npy respectively. V is the victim class of the expirements (out of ModelNet 40 classes ) and T is the target class (100 if untargeted attack) , and B is the batch number. By default the code will iterate over all the victims and targets in our test data data/attacked_data.z. A summary table of the evaluation of teh attack output will be saved in results/exp0/exp0_all.csv

Other files

  • log/NETWORK/model.ckpt -- the victims models (trained on ModelNet40) used in the paper.
  • 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)
  • 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 aligned version of ModelNet40 data (in point cloud data format) can be downloaded here.
  • The visulization in the paper is rendered with pptk
  • Please open an issue or contact Abdullah Hamdi (abdullah.hamdi@kaust.edu.sa) if there is any question.

Acknoledgements

This paper and repo borrows codes and ideas from several great github repos: latent 3D point clouds , 3d-adv-pc, Dynamic Graph CNN for Learning on Point Clouds, PointNet ++

License

The code is released under MIT License (see LICENSE file for details).