Prototype-supervised Adversarial Network for Targeted Attack of Deep Hashing

This is the code for our CVPR 2021 paper "Prototype-supervised Adversarial Network for Targeted Attack of Deep Hashing", which formulates a flexible generative architecture for efficient and effective targeted hashing attack. In this repository, we not only provide the implementation of the proposed Prototype-supervised Adversarial Network (i.e., ProS-GAN), but also collect some popular deep hashing methods used in the paper and the previous targeted attack methods in hashing based retrieval.

The jounal version extension of this paper has been published on IEEE TMM 2022.

Usage

Dependencies

• Python 3.7.6
• Pytorch 1.6.0
• Numpy 1.18.5
• Pillow 7.1.2
• CUDA 10.2

Train hashing models

Initialize the hyper-parameters in hashing.py following the paper, and then run

python hashing.py

Attack by P2P or DHTA

Initialize the hyper-parameters in dhta.py following the paper, and then run

python dhta.py

Train ProS-GAN

Initialize the hyper-parameters in main.py following the paper, and then run

python main.py --train True

Evaluate ProS-GAN

Initialize the hyper-parameters in main.py following the paper, and then run

python main.py --train False --test True

Cite

If you find this work is useful, please cite the following:

@inproceedings{wang2021prototype,
	title={Prototype-supervised Adversarial Network for Targeted Attack of Deep Hashing},
	author={Wang, Xunguang and Zhang, Zheng and Wu, Baoyuan and Shen, Fumin and Lu, Guangming},
	booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
	year={2021}
}