LogoStyleFool

This is the source code for our paper "LogoStyleFool: Vitiating Video Recognition Systems via Logo Style Transfer" (AAAI 2024).

Requirements

• python == 3.6
• pytorch == 1.10.0
• kornia == 0.2.2
• torchvision == 0.11.3
• easydict
• opencv
• scikit-learn
• tqdm
• scipy

Dataset

Please download the action recognition dataset UCF-101 and HMDB51, then process and save them in 'data/'. We use the same preprocessing as StyleFool.

Pretrained model

The pre-trained model for C3D on UCF-101, as well as models for style transfer, is provided here.

Usage

LogoStyleFool

Targeted attack

Run python main.py --model C3D --dataset UCF101 --video_npy_path ./your/path/BenchPress/v_BenchPress_g20_c06.npy --label 9 --target --target_class 55 --output_path result/.

Untargeted attack

Run python main.py --model C3D --dataset UCF101 --video_npy_path ./your/path/FrontCrawl/v_FrontCrawl_g09_c03.npy --label 31 --output_path result/.

Basic arguments:

• --model: The attacked model.
• --dataset: The dataset.
• --gpu: ID of the GPU to use.
• --video_npy_path: The video path in npy forms.
• --label: The label of the video.
• --target: Targeted attack or untargeted attack (default).
• --target_class: Targeted attack class.
• --output_path: The path to save output_adversarial_npy_path.
• --rl_batch: The batch size of RL.
• --steps: The steps of RL.
• --sigma: The RL reward ratio to control area.
• --tau: The RL reward ratio to control distance.
• --logo_num: The num of logos.
• --style_num: The num of style imgs.
• --max_iters: The max iters of LogoS-DCT.
• --epsilon: The epsilon of LogoS-DCT.
• --linf_bound: The linf bound of perturbation.(0 ~ 1 for LogoStyleFool-$l_2$ and 0 for LogoStyleFool-$l_\infty$)

Acknowledgement

• Part of our implementation is based on LinearStyleTransfer, PatchAttack and simple-blackbox-attack. We thank for their extraordinary contributions.

Citation

If you use this code or its parts in your research, please cite the following paper:

@inproceedings{cao2024logostylefool,
      title={LogoStyleFool: Vitiating Video Recognition Systems via Logo Style Transfer}, 
      author={Cao, Yuxin and Zhao, Ziyu and Xiao, Xi and Wang, Derui and Xue, Minhui and Lu, Jin},
      booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
      year={2024},
      address={Vancouver, Canada},
      month={February}
}