CT-GAT

Code, data and model parameter of the EMNLP 2023 paper "CT-GAT: Cross-Task Generative Adversarial Attack based on Transferability"

This repository contains code modified from Advbench, many thanks!

Dependencies

You can create the same running environment and install dependencies as us by using the following commands:

conda env create -f environment.yaml

Data Preparation and Preprocess

Please visit the /data folder, read the README.md there to obtain and process the data.

Folder Creation

Execute the following code to create the required folders

mkdir param
mkdir victim
mkdir output

Experiments

In this step, you need to operate under the CT-GAT directory.

First, you need to train the CT-GAT generator. You can run the following command for training. You can also directly download our parameters from Google Cloud: here. Or you can download our trained model parameters from Baidu Cloud: here

bash scripts/train_CT-GAT.sh

Then you should fine-tune the pre-trained model on our security datasets collection Advbench.

bash scripts/train_victim.sh

To conduct the baseline attack experiments in default settings:

bash scripts/base_attack.sh

To conduct attack experiments via ROCKET in default settings:

bash scripts/ROCKT.sh

To conduct attack experiments via CT-GAT in our settings:

bash scripts/CT-GAT.sh

Citation

Please kindly cite our paper:

@misc{lv2023ctgat,
      title={CT-GAT: Cross-Task Generative Adversarial Attack based on Transferability}, 
      author={Minxuan Lv and Chengwei Dai and Kun Li and Wei Zhou and Songlin Hu},
      year={2023},
      eprint={2310.14265},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}