tta-gan's Introduction

Targeted Transferable Attack against Deep Hashing Retrieval

This is the code for "Targeted Transferable Attack against Deep Hashing Retrieval".

Requirements

python 3.8
torch 1.10
torchvision 0.11
numpy 1.20

Implementation

Overview of previous methods

We consider four previous targeted attack methods as our competitors: DHTA (P2P), THA, ProS_GAN and NAG.
To evaluate the targeted transferability of different methods, we test the performance on five deep hashing methods, i.e., DPSH, HashNet, CSQ, DSDH and ADSH. Our implementations are modified based on DeepHash-pytorch.

Overview of dirs and files

data contains the dataset files utilized in this paper.
Hash contains the implementations of five deep hashing methods.
utils contains all the tools used for training models.
models contains the implementations of our method.

Usage

Train deep hashing models

You can easily train deep hashing models by replacing the path of data in the code, and then run

cd Hash
python DPSH.py
python HashNet.py
python CSQ.py
python DSDH.py
python ADSH.py

Generate anchor code

After setting the dataset and target model paths, you can generate anchor code by running

cd models
python IAO.py

Attack by our TTA-GAN method

Initialize the hyperparameters following our paper and then run

cd models
python TTA_GAN.py

Ensemble attack

To conduct model ensemble attack, you can run

cd models
python Ens.py

Acknowledgement

The codes are modified based on Wang et al. 2021.

Cite

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

@inproceedings{zhu2023targeted,
  title={Targeted Transferable Attack against Deep Hashing Retrieval},
  author={Zhu, Fei and Zhang, Wanqian and Wu, Dayan and Wang, Lin and Li, Bo and Wang, Weiping},
  booktitle={Proceedings of the 5th ACM International Conference on Multimedia in Asia},
  pages={1--7},
  year={2023}
}

Recommend Projects