Skip to content

Official implementation for "Consistent Attack: Universal Adversarial Perturbation on Embodied Vision Navigation" (PRL 2023)

License

Notifications You must be signed in to change notification settings

yingchengyang/Consistent-Attack

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

16 Commits
 
 
 
 
 
 
 
 

Repository files navigation

Consistent-Attack

arXiv

Installation

git clone this repo

cd Consistent-Attack
conda create -n consistent-attack python=3.7
conda activate consistent-attack

Then you need to install habitat-sim and then

pip install -e habitat-lab
pip install -e habitat-baselines
pip install numpy
pip install torch

Usage

Then you can run the attack code like

cd habitat-baselines/habitat-baselines

xvfb-run -a python run.py --exp-config config/pointnav/ppo_pointnav_test_d.yaml --run-type eval habitat_baselines.torch_gpu_id 0 habitat_baselines.simulator_gpu_id 0 habitat_baselines.eval.evaluate_strategy 1 habitat_baselines.eval.update_num 5 habitat_baselines.eval.traj_num_each 1

here we do not attack the model by setting habitat_baselines.eval.evaluate_stratege=0, use UAP by setting habitat_baselines.eval.evaluate_stratege=1, use Reward UAP by setting habitat_baselines.eval.evaluate_stratege=2, use Trajectory UAP by setting habitat_baselines.eval.evaluate_stratege=3

Pretrained models are in the fold "/habitat-baselines/model/" and are from Habitat-baselines in Habitat-Lab.

Scene datasets are in the fold "/habitat-baselines/scene_data/". We only provide scene datasets of habitat-test from Habitat-Lab. The dataset of gibson and mp3d can be downloaded by the instructions here.

Citation

If you find Consistent Attack helpful, please cite our paper.

@article{ying2023consistent,
  title={Consistent Attack: Universal Adversarial Perturbation on Embodied Vision Navigation},
  author={Ying, Chengyang and Qiaoben, You and Zhou, Xinning and Su, Hang and Ding, Wenbo and Ai, Jianyong},
  journal={Pattern Recognition Letters},
  year={2023},
  publisher={Elsevier}
}

About

Official implementation for "Consistent Attack: Universal Adversarial Perturbation on Embodied Vision Navigation" (PRL 2023)

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published