README.md

llm_defends

code of paper "Defending Against Alignment-Breaking Attacks via Robustly Aligned LLM"

Environment Setup

First, our code requires the environment listed in requirements.txt. to be installed:

pip install -r requirements.txt

download LLM models

Before running the code, you need to first download the model weights locally: vicuna model weights, guanaco model weights.

For the LLaMA model, we use the Llama-2-7b-chat-hf model, please apply for it on the official website and download it.

After downloading the models, please modify the model_dirs variable in the defaults.py file to correspond with your model storage path.

Run Experiments

For security and to prevent abuse, we only provide three adversarial prompts for the vicuna model to run experiment demos.

To run the experiment demo, please execute the following command:

python main.py  --sampling_num=20 \
                --max_new_tokens=10 \
                --dropping_rate=0.3 \
                --padding=True \
                --test_model=vicuna \
                --threshold=0.2 \
                --manual_seed=0 \
                --device=0 \

Below are detailed explanations of the hyperparameters:

• sampling_num: Number of Monte Carlo trials, corresponding to $n$ in the paper, default is 20.
• max_new_tokens: Maximum token number generated by LLM, corresponding to $t_{max}$ in the paper, default is 10.
• dropping_rate: Proportion of tokens randomly dropped during each Monte Carlo trial, corresponding to $p$ in the paper, default is 0.3.
• padding: Whether to perform padding on the input. For the Vicuna model, set it as True, for the LLaMA and Guanaco models, set it as False.
• test_model: The name of the model used for testing, can be vicuna, llama, or guanaco.
• threshold: Threshold used to determine whether a query is inappropriate, corresponding to $t$ in the paper, default is 0.2.
• manual_seed: Random seed, default is 0.
• device: GPU number to be used for running the code, default is 0.

Bibtex

If these codes have been helpful to you, welcome to cite our paper! The bibtex of our paper is:

@misc{cao2023defending,
      title={Defending Against Alignment-Breaking Attacks via Robustly Aligned LLM}, 
      author={Bochuan Cao and Yuanpu Cao and Lu Lin and Jinghui Chen},
      year={2023},
      eprint={2309.14348},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}