Skip to content

Latest commit

 

History

History
79 lines (60 loc) · 3.46 KB

README.md

File metadata and controls

79 lines (60 loc) · 3.46 KB

Adversarial attacks against Stereo R-CNN

This folder contains code about performing perturbation attack and patch attack against one of our target vision-based 3D object detectors, namely, Stereo-RCNN.

Requirement

This implementation is tested in following environment:

The perturbation and patch are trained with 1 NVIDIA Tesla T4 GPU.

Install

Please refer instructions in Stereo R-CNN 0.Install to install the repo in two folder, ./Stereo R-CNN and ./eval/Stereo R-CNN, for attack and eval respectively.

Third party libraries can be installed(in a python3 virtualenv) using:

pip install -r requirements.txt

Dataset preparation

Please follow steps in the Dataset Preparation to set the dataset. And use symbolic links to make data/kitti/object/training/image_2 and data/kitti/object/training/image_3 changeable. So that, they can point to image folder we generated under attacks.

After setting up the repo, please add our attack code to two directories. Remember to substitute stereo_rpn.py, stereo_rcnn.py and roibatchLoader.py under attack repo. The overall directory will look like below:

.                                           (root directory)
|-- Stereo R-CNN
    |-- lib
        |-- model
            |-- stereo_rpn.py               (substitute file)
            |-- stereo_rcnn.py              (substitute file)
            |-- ...
        |-- roi_data_layer
            |-- roibatchLoader.py           (substitute file)
            |-- ...
    |-- patch_attack.py                     (patch attack file)
    |-- pgd_attack.py                       (pgd attack file)
    |-- ...            
|-- DSGN
    |-- predict_and_save_patch.py           (evaluate patch attack)
    |-- predict_and_save_pgd.py             (evaluate pgd attack)
    |-- ...

Perturbation attack

Launch the perturbation attack:

python pgd_attack.py --iter [iteration number of attack] --alpha [perturbation update step of attack]

The attacked images are saved in ./stereo_rcnn_pgd_iters_[iteration number] folder.

Evaluate Stereo R-CNN under perturbation attack (please ensure that you have change the symbolic links of data/kitti/object/training/image_2 and data/kitti/object/training/image_3 to the image folder under attack):

python predict_and_save_pgd.py --iter [iteration number of attack] --alpha [perturbation update step of attack]

The detection results are saved in result_stereo_rcnn_pgd_[iteration number]_{alpha value}. If you want to evaluate regarding average precision, please refer to kitti_eval.

Patch attack

Launch the patch attack:

python patch_attack.py --debug --debugnum [number of training data] --epochs [training epoch number] --ratio [trained patch size]

The trained patch is saved in ./stereo_rcnn_patch_ratio_[ratio size] folder.

Evaluate Stereo R-CNN under patch attack (please ensure the symbolic links of image_2 and image_3 point to raw data):

python predict_and_save_patch.py -ratio [trained patch ratio] --epochs [training epoch number] --patch_dir [path to your patch folder] --atk_mode [random/specific attack]