You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
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.
About
Official pytorch implementation of NeurIPS 2021 paper Geo-TRAP