Geo-TRAP

Official pytorch implementation of NeurIPS 2021 paper Adversarial Attacks on Black Box Video Classifiers: Leveraging the Power of Geometric Transformations.

Installation

See installation instruction here.

Prepare data

For details on data preparation, you can refer to

• preparing_ucf101
• preparing_jester

Pre-trained video models

The pretrained C3D, SlowFast, TPN and I3D model on both UCF-101 and Jester dataset can be found in Dropbox.

Usage

Here, we give an example of how to do targeted attack to C3D model on Jester dataset with affine transformation.

python query_attack/decompose_query.py 
    --targeted # targeted attack 
    --config configs/recognition/c3d/c3d_jester_MotionPerturbation.py  # contains configuration of Jester dataset                         
    --config_rec configs/recognition/c3d/c3d_sports1m_16x1x1_45e_jester_rgb.py # contains configuration of C3D model                         
    --checkpoint_rec work_dirs/c3d_sports1m_16x1x1_45e_jester_rgb/epoch_30.pth # pretrained C3D model paramters                                            
    --transform_type_query  affine # use affine geometric transformation

For untargeted attacks, other video models (SlowFast, TPN and I3D) and UCF101 dataset, please see adversarial attack commands here

Citation

If you find our work helpful in your research, please cite it as

@article{li2021adversarial,
  title={Adversarial Attacks on Black Box Video Classifiers: Leveraging the Power of Geometric Transformations},
  author={Li, Shasha and Aich, Abhishek and Zhu, Shitong and Asif, M Salman and Song, Chengyu and Roy-Chowdhury, Amit K and Krishnamurthy, Srikanth},
  journal={arXiv preprint arXiv:2110.01823},
  year={2021}
}

Acknowledgement

Many thanks to MMAction2 for the video model implementation.