论文 GCG 复现 Universal and Transferable Adversarial Attacks on Aligned Language Models

0 前言

总目录 大模型安全相关研究：https://blog.csdn.net/WhiffeYF/article/details/142132328

https://github.com/llm-attacks/llm-attacks

https://github.com/lm-sys/FastChat

论文翻译：Universal and Transferable Adversarial Attacks on Aligned Language Models

b站视频：https://www.bilibili.com/video/BV1R553zkE92/

平台采用Autodl：https://www.autodl.com/home

PyTorch / 2.3.0 / 3.12(ubuntu22.04) / 12.1
vGPU-48GB

1 模型下载

Llama-2-7b-chat-hf 模型下载

使用魔塔社区下载：https://www.modelscope.cn/models/shakechen/Llama-2-7b-chat-hf/files

使用SDK下载下载：

开始前安装

source /etc/network_turbo
pip install modelscope

脚本下载

# source /etc/network_turbo
from modelscope import snapshot_download

# 指定模型的下载路径
cache_dir = '/root/autodl-tmp'
# 调用 snapshot_download 函数下载模型
model_dir = snapshot_download('shakechen/Llama-2-7b-chat-hf', cache_dir=cache_dir)

print(f"模型已下载到: {model_dir}")

2 llm-attacks 安装

source /etc/network_turbo

git clone https://github.com/llm-attacks/llm-attacks.git

requirements.txt需要做修改：

https://github.com/llm-attacks/llm-attacks/blob/main/requirements.txt

transformers==4.28.1
ml_collections
fschat==0.2.20

改成

transformers

接着继续安装

cd llm-attacks

pip install -e .

source /etc/network_turbo
pip install ml_collections

pip install pandas

加载模型位置修改：
https://github.com/llm-attacks/llm-attacks/blob/main/experiments/configs/individual_llama2.py

路径修改为：/root/autodl-tmp/shakechen/Llama-2-7b-chat-hf

3 FastChat 安装

source /etc/network_turbo

git clone https://github.com/lm-sys/FastChat.git
cd FastChat

pip install -e ".[model_worker,webui]"

4 Experiments

cd experiments/launch_scripts
bash run_gcg_individual.sh llama2 behaviors

5 结果