:see_no_evil: U-Turn :hear_no_evil:

Attack your retrieval model via Query! They are not robust as you expected!

One simple code to cheat your retrieval model via Modifying Query ONLY (based on pytorch) accepted by IJCV. Pre-print version is at https://arxiv.org/abs/1809.02681.

The main idea underpinning our method is simple yet effective, making the query feature to conduct a U-turn :arrow_right_hook:.

Table of contents

• Re-ID Attacking
• Image Retrieval Attacking
• Cifar Attacking

Re-ID Attacking

1.1 Preparing your reID models.

Please check the step-by-step tutorial in https://github.com/layumi/Person_reID_baseline_pytorch

1.2 Attacking Market-1501

Try four attack methods with one line. Please change the path before run it.

python experiment.py

Image Retrieval Attacking

2.1 Download the pre-trained model on Oxford and Paris

We attach the training code, which is based on the excellent code in TPAMI 2018. https://github.com/layumi/Oxford-Paris-Attack

2.2 Attacking the Oxford and Paris Dataset

Our effort is to cheat the TPAMI model. Yes. We succeed. https://github.com/layumi/Oxford-Paris-Attack

2.3 Attacking Food-256 and CUB-200-2011

Please check subfolders.

Food: https://github.com/layumi/U_turn/tree/master/Food

CUB: https://github.com/layumi/U_turn/tree/master/cub

Cifar Attacking

3.1 Cifar (ResNet-Wide)

We attach the training code, which is borrowed from ResNet-Wide (with Random Erasing).

3.2 Attacking Cifar

https://github.com/layumi/A_reID/tree/master/cifar

Citation

@article{zheng2022query,
  title={U-turn: Crafting Adversarial Queries with Opposite-direction Features},
  author={Zheng, Zhedong and Zheng, Liang and Yang, Yi and Wu, Fei},
  journal={IJCV},
  year={2022}
}