DSN Jailbreak Attack & Evaluation Ensemble

This is the official repository for Don't Say No: Jailbreaking LLM by Suppressing Refusal

In this work, we propose one novel jailbreak optmization target, introducing the refusal loss, cosine decay weighting schedule method, applying the Unlikelihood loss to ensure convergence, and novelly incorporate a NLI based method into our proposed jailbreak Evaluation Ensemble pipeline.

Installation

Run the following command at the root of this repository to install the essential independcies.

pip install -e .
python -m spacy download en_core_web_sm

Note that we have chosen a different transformers version from the default GCG implementation, which might introduce subtle difference in Llama-2 model conversation prompt formatting. To faithfully reproduce all the results reported in the paper, e.g. both DSN and GCG attack results upon Llama-2 model, try install the transformers package with version 4.28.1 and fschat package with version 0.2.20. Both version of conda environmnets are supported by our implementation.

To reproduce

First, download all the target models as well as the NLI and $\textit{HarmBench}$ model utilized in Evaluation Ensemble, and configure the OpenAI client in file eval_ensemble_src/eval2_gpt.py properly.

The attack launch file is located as experiments/launch_scripts/run_dsn_attack.sh

The recommend hyper-parameter setting is already detailed in the script. For rigorous evaluation, please consider running the same setting for multiple times.

To evaluate, please first run the scripts in eval_scripts folder, e.g. experiments/eval_scripts/lastStep_eval.sh, to conduct evaluation under the Refusal Matching metric. Then, change directory to eval_ensemble_src to launch eval_ensemble_src/launch_majority_eval.sh

Updates

• 2024.11.6: Refine the _update_ids function, to support llama-2 conv_template within different transformers and fschat version. For more details please refer to Section B.1.1 of our paper

Citation

If you find our implementation and paper useful, please consider citing our work:

@misc{zhou2024dontsaynojailbreaking,
      title={Don't Say No: Jailbreaking LLM by Suppressing Refusal}, 
      author={Yukai Zhou and Wenjie Wang},
      year={2024},
      eprint={2404.16369},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2404.16369}, 
}

In addition, please also consider citing the following works, which have been utilized in this paper

@misc{zou2023universal,
  title={Universal and Transferable Adversarial Attacks on Aligned Language Models},
  author={Andy Zou and Zifan Wang and J. Zico Kolter and Matt Fredrikson},
  year={2023},
  eprint={2307.15043},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}

@article{mazeika2024harmbench,
  title={HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal},
  author={Mazeika, Mantas and Phan, Long and Yin, Xuwang and Zou, Andy and Wang, Zifan and Mu, Norman and Sakhaee, Elham and Li, Nathaniel and Basart, Steven and Li, Bo and Forsyth, David and Hendrycks, Dan},
  journal={arXiv preprint arXiv:2402.04249},
  year={2024}
}

@inproceedings{chao2024jailbreakbench,
  title={JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models},
  author={Patrick Chao and Edoardo Debenedetti and Alexander Robey and Maksym Andriushchenko and Francesco Croce and Vikash Sehwag and Edgar Dobriban and Nicolas Flammarion and George J. Pappas and Florian Tramèr and Hamed Hassani and Eric Wong},
  booktitle={NeurIPS Datasets and Benchmarks Track},
  year={2024}
}

@article{huang2023catastrophic,
  title={Catastrophic Jailbreak of Open-source LLMs via Exploiting Generation},
  author={Huang, Yangsibo and Gupta, Samyak and Xia, Mengzhou and Li, Kai and Chen, Danqi},
  journal={arXiv preprint arXiv:2310.06987},
  year={2023}
}

Acknowledgment

DSN attack implementation is built upon modifying the GCG implementaion from GCG github repo

Sincere thanks to the authors of Zou et al. 2023 for the valueable insight and open-sourced code.

License

DSN is licensed under the terms of the MIT license. See LICENSE for more details.