Adversarial-Prompt-Translator

This repository contains a PyTorch implementation for our paper Deciphering the Chaos: Enhancing Jailbreak Attacks via Adversarial Prompt Translation.

Environments

• Python 3.8.8
• PyTorch 2.4.0
• transformers 4.44.2
• vllm 0.5.4
• openai 1.44.1

Usage

To generate adversarial prompts using Llama-3.1-8B as the translator LLM on the HarmBench dataset, run:

translate.py --translator llama3.1-8b --dataset harmbench

The adversarial prompts will be saved in results/trans_harmbench_llama3.1-8b.json.

For evaluation, we use OpenAI's Batch API. To evaluate the performance of attacking a GPT/Claude model, for instance, gpt-4o-mini, submit the adversarial prompts as a batch by running:

logpath=results/trans_harmbench_llama3.1-8b.json model=gpt-4o-mini bash eval_submit.sh

You can track the status of the generation by running:

python3 eval_openai.py --v

The output will be like:

id   status        description
0    in_progress   results/trans_harmbench_llama3.1-8b_eval_gpt-4o-mini.jsonl
...  ...           ...

The "id" is used for identifying different submissions. Once the status turns to "completed", download the output file by running:

python3 eval_openai.py --d {id}

The output file will be saved as results/trans_harmbench_llama3.1-8b_eval_gpt-4o-mini_output.jsonl. Then, to evaluate the attack success rate using the classifier of HarmBench, run:

bash eval_harmbench.sh results/trans_harmbench_llama3.1-8b_eval_gpt-4o-mini_output.jsonl

The results will be saved in results/trans_harmbench_llama3.1-8b_eval_gpt-4o-mini_eval.json

Citation

Please cite our work in your publications if it helps your research:

@article{li2024deciphering,
         title={Deciphering the Chaos: Enhancing Jailbreak Attacks via Adversarial Prompt Translation},
         author={Li, Qizhang and Yang, Xiaochen and Zuo, Wangmeng and Guo, Yiwen},
         journal={arXiv preprint arXiv:2410.11317},
         year={2024}
}