DGSlow

Codebase for the ACL 2023 paper: "White-Box Multi-Objective Adversarial Attack on Dialogue Generation" (PDF).

Quickstart

Setup Environment

• python 3.10.8
• pytorch 1.13.0+
• Install dependencies

pip install -r requirements.txt

Train and evaluate a model on a specific task(s)

• BART in Blended Skill Talk:

python train_seq2seq.py --model_name_or_path facebook/bart-base --dataset blended_skill_talk --output_dir results/bart-base

• DialoGPT in Empathetic Dialogues:

python train_clm.py --model_name_or_path microsoft/DialoGPT-small --dataset empathetic_dialogues --output_dir results/dialogpt-small

Attack a pre-trained model

• Structure attack on DialoGPT-small in Blended Skill Talk:

python attack.py --attack_strategy structure --model_name_or_path results/bart-base --dataset blended_skill_talk

• DF attack on bart-base in Empathetic Dialogues:

python attack.py --attack_strategy FD --model_name_or_path results/bart-base --dataset empathetic_dialogues

Transfer attack

• Transfer attack from DialoGPT-small to bart-base in Blended Skill Talk:

python eval.py --file ${FILE} --orig_model bart-base --victim_model dialogpt-small --dataset BST --out_dir logging

Citation

Please cite the paper in your publications if you find this repo useful:

@inproceedings{li2023white,
  title={White-Box Multi-Objective Adversarial Attack on Dialogue Generation},
  author={Li, Yufei and Li, Zexin and Gao, Yingfan and Liu, Cong},
  booktitle={Annual Meeting of the Association for Computational Linguistics (ACL)},
  year={2023}
}

Acknowledgement

Our implementation is based on OpenAttack. We would like to thank the authors for their open source code.