BESA: : BERT-based Simulated Annealing for Adversarial Text Attacks (IJCAI21)

This repository contains the implementations of the IJCAI-21 paper: BESA: BERT-based Simulated Annealing for Adversarial Text Attacks.

Usage

Our code can be easily tested by the command line like:

textattack attack --recipe bert-sa --model bert-base-uncased-imdb --num-examples 1000

This will attack BERT model on IMDB dataset using our BESA.

Citation

When using this code, or the ideas of BESA, please cite the following paper

@inproceedings{yang2021besa,
  title={BESA: BERT-based Simulated Annealing for Adversarial Text Attacks},
  author={Xinghao Yang and Weifeng Liu and Dacheng Tao and Wei Liu},
  booktitle={Proceedings of the 30th International Joint Conference on Artificial Intelligence},
  year={2021}
}

Acknowledgement

This code is based on the TextAttack framework.

Contact

Please contact Xinghao Yang or Wei Liu at firstname.lastname@uts.edu.au if you're interested to collaborate on this research!