Learning to Attack Real-World Models for Person Re-identification via Virtual-Guided Meta-Learning (AAAI 2021)

Code for AAAI 2021 paper ``Learning to Attack Real-World Models for Person Re-identification via Virtual-Guided Meta-Learning (AAAI 2021)".

Paper

Supplemental Material

Requirements:

• python 3.7
• CUDA==10.1
• Market1501, DukeMTMC-reID, MSMT-17 and PersonX456.
• faiss-gpu==1.6.0
• Other necessary packages listed in requirements.txt

Preparing Data

• Clone our repo

Market-1501 (Duke and MSMT are the same):

• Download "Market-1501-v15.09.15.zip".
• Create a new directory, rename it as "data".
• Create a directory called "raw" under "data" and put "Market-1501-v15.09.15.zip" under it.
• The final structure should like this (meta.json and splits.json are generated by our code, please ignore them):

PersonX:

• Download full PersonX dataset from https://github.com/sxzrt/Instructions-of-the-PersonX-dataset, we use the CVPR'19 version.
• Merge "query" of "4", "5", "6" together (the same to "bounding_box_train" and "bounding_box_test").
• Put the merged "query", "bounding_box_train" and "bounding_box_test" to a new directory "personX456", zip it and put it to data/personx/raw
• The final structure should like this:

• There is a processed tar file in BaiduYun with all needed files, password: 4fqf. You can directly put it under "data" and unzip it.

Preparing Models

• Download re-ID models from BaiduYun (Password: tua2)

• Put models under logs/{datasetname}

Run our code

See attack.sh for more information.

If you find this code useful in your research, please consider citing:

@inproceedings{yang2021learning,
  title={Learning to Attack Real-World Models for Person Re-identification via Virtual-Guided Meta-Learning},
  author={Yang, Fengxiang and Zhong, Zhun and Liu, Hong and Wang, Zheng and Luo, Zhiming and Li, Shaozi and Sebe, Nicu and Satoh, Shin’ichi},
  booktitle={AAAI},
  volume={35},
  number={4},
  pages={3128--3135},
  year={2021}
}

Acknowledgments

Our code is based on UAP-Retrieval, if you use our code, please also cite their paper.

@inproceedings{Li_2019_ICCV,
    author = {Li, Jie and Ji, Rongrong and Liu, Hong and Hong, Xiaopeng and Gao, Yue and Tian, Qi},
    title = {Universal Perturbation Attack Against Image Retrieval},
    booktitle = {ICCV},
    year = {2019}
}

Contact Me

Email: yangfx@stu.xmu.edu.cn