Skip to content

A framework to evaluate the generalization capability of safety alignment for LLMs

License

Notifications You must be signed in to change notification settings

RobustNLP/CipherChat

Repository files navigation

CipherChat 🔐

A novel framework CipherChat to systematically examine the generalizability of safety alignment to non-natural languages – ciphers.

If you have any questions, please feel free to email the first author: Youliang Yuan.

👉 Paper

For more details, please refer to our paper ICLR 2024.

Logo

LOVE💗 and Peace🌊

RESEARCH USE ONLY✅ NO MISUSE❌

Our results

We provide our results (query-response pairs) in experimental_results, these files can be loaded by torch.load(). Then, you can get a list: the first element is the config and the rest of the elements are the query-response pairs.

result_data = torch.load(filename)
config = result_data[0]
pairs = result_data[1:]

🛠️ Usage

✨An example run:

python3 main.py \
 --model_name gpt-4-0613 \
--data_path data/data_en_zh.dict \
--encode_method caesar \
--instruction_type Crimes_And_Illegal_Activities \
--demonstration_toxicity toxic \
--language en

🔧 Argument Specification

  1. --model_name: The name of the model to evaluate.

  2. --data_path: Select the data to run.

  3. --encode_method: Select the cipher to use.

  4. --instruction_type: Select the domain of data.

  5. --demonstration_toxicity: Select the toxic or safe demonstrations.

  6. --language: Select the language of the data.

💡Framework

Logo

Our approach presumes that since human feedback and safety alignments are presented in natural language, using a human-unreadable cipher can potentially bypass the safety alignments effectively. Intuitively, we first teach the LLM to comprehend the cipher clearly by designating the LLM as a cipher expert, and elucidating the rules of enciphering and deciphering, supplemented with several demonstrations. We then convert the input into a cipher, which is less likely to be covered by the safety alignment of LLMs, before feeding it to the LLMs. We finally employ a rule-based decrypter to convert the model output from a cipher format into the natural language form.

📃Results

The query-responses pairs in our experiments are all stored in the form of a list in the "experimental_results" folder, and torch.load() can be used to load data.

Logo

🌰Case Study

Logo

🫠Ablation Study

Logo

🦙Other Models

Logo

Star History Chart

Community Discussion:

Citation

If you find our paper&tool interesting and useful, please feel free to give us a star and cite us through:

@inproceedings{
yuan2024cipherchat,
title={{GPT}-4 Is Too Smart To Be Safe: Stealthy Chat with {LLM}s via Cipher},
author={Youliang Yuan and Wenxiang Jiao and Wenxuan Wang and Jen-tse Huang and Pinjia He and Shuming Shi and Zhaopeng Tu},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=MbfAK4s61A}
}

Releases

No releases published

Packages

No packages published

Languages