cgat's Introduction

CgAT: Center-Guided Adversarial Training for Deep Hashing-Based Retrieval

This is the official code for CgAT: Center-Guided Adversarial Training for Deep Hashing-Based Retrieval (WWW 2023).

In this paper, we present a min-max based Center-guided Adversarial Training, namely CgAT, to improve the robustness of deep hashing networks through worst adversarial examples. Our key idea is to formulate a hash code (dubbed center code) as a discriminative semantic representation of the original sample, which can be used to guide the generation of the powerful adversarial example and as an accurate optimization objective for adversarial training.

Requirements

My dependencies are as follows:

python 3.9.12
pytorch 1.12.0
torchvision 0.13.0
pillow 9.0.1
numpy 1.21.5
cudatoolkit 11.3.1

Quick Start

Train hashing models

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

python hashing.py

Adversarial training

Initialize the hyper-parameters in central_adv_train.py according to the paper, and then run

python central_adv_train.py

Citation

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

@inproceedings{wang2023cgat,
  title={CgAT: Center-Guided Adversarial Training for Deep Hashing-Based Retrieval},
  author={Wang, Xunguang and Lin, Yiqun and Li, Xiaomeng},
  booktitle={Proceedings of the ACM Web Conference (WWW)},
  year={2023}
}

cgat's People

Contributors

Stargazers

Watchers

cgat's Issues

How to reproduce the results.

What is the difference between this method and SAAT method. Besides, i wanna know how to reproduce the results reported in the paper and it seems that the robust model can not be trained successfully. Please answer this question for me, thanks a lot.