# llm-attacks

**Repository Path**: citiao/llm-attacks

## Basic Information

- **Project Name**: llm-attacks
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: MIT
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2024-07-31
- **Last Updated**: 2024-07-31

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# LLM Attacks

[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)

This is the official repository for "[Universal and Transferable Adversarial Attacks on Aligned Language Models](https://arxiv.org/abs/2307.15043)" by [Andy Zou](https://andyzoujm.github.io/), [Zifan Wang](https://sites.google.com/west.cmu.edu/zifan-wang/home), [Nicholas Carlini](https://nicholas.carlini.com/), [Milad Nasr](https://people.cs.umass.edu/~milad/), [J. Zico Kolter](https://zicokolter.com/), and [Matt Fredrikson](https://www.cs.cmu.edu/~mfredrik/).

Check out our [website and demo here](https://llm-attacks.org/).

## Updates
- (2023-08-16) We include a notebook `demo.ipynb` (or see it on [Colab](https://colab.research.google.com/drive/1dinZSyP1E4KokSLPcCh1JQFUFsN-WV--?usp=sharing)) containing the minimal implementation of GCG for jailbreaking LLaMA-2 for generating harmful completion.


## Table of Contents

- [Installation](#installation)
- [Models](#models)
- [Experiments](#experiments)
- [Demo](#demo)
- [Reproducibility](#reproducibility)
- [License](#license)
- [Citation](#citation)

## Installation

We need the newest version of FastChat `fschat==0.2.23` and please make sure to install this version. The `llm-attacks` package can be installed by running the following command at the root of this repository:

```bash
pip install -e .
```

## Models

Please follow the instructions to download Vicuna-7B or/and LLaMA-2-7B-Chat first (we use the weights converted by HuggingFace [here](https://huggingface.co/meta-llama/Llama-2-7b-hf)).  Our script by default assumes models are stored in a root directory named as `/DIR`. To modify the paths to your models and tokenizers, please add the following lines in `experiments/configs/individual_xxx.py` (for individual experiment) and `experiments/configs/transfer_xxx.py` (for multiple behaviors or transfer experiment). An example is given as follows.

```python
    config.model_paths = [
        "/DIR/vicuna/vicuna-7b-v1.3",
        ... # more models
    ]
    config.tokenizer_paths = [
        "/DIR/vicuna/vicuna-7b-v1.3",
        ... # more tokenizers
    ]
```

## Demo
We include a notebook `demo.ipynb` which provides an example on attacking LLaMA-2 with GCG. You can also view this notebook on [Colab](https://colab.research.google.com/drive/1dinZSyP1E4KokSLPcCh1JQFUFsN-WV--?usp=sharing). This notebook uses a minimal implementation of GCG so it should be only used to get familiar with the attack algorithm. For running experiments with more behaviors, please check Section Experiments. To monitor the loss in the demo we use `livelossplot`, so one should install this library first by pip.

```bash
pip install livelossplot
```

## Experiments 

The `experiments` folder contains code to reproduce GCG experiments on AdvBench.

- To run individual experiments with harmful behaviors and harmful strings (i.e. 1 behavior, 1 model or 1 string, 1 model), run the following code inside `experiments` (changing `vicuna` to `llama2` and changing `behaviors` to `strings` will switch to different experiment setups):

```bash
cd launch_scripts
bash run_gcg_individual.sh vicuna behaviors
```

- To perform multiple behaviors experiments (i.e. 25 behaviors, 1 model), run the following code inside `experiments`:

```bash
cd launch_scripts
bash run_gcg_multiple.sh vicuna # or llama2
```

- To perform transfer experiments (i.e. 25 behaviors, 2 models), run the following code inside `experiments`:

```bash
cd launch_scripts
bash run_gcg_transfer.sh vicuna 2 # or vicuna_guanaco 4
```

- To perform evaluation experiments, please follow the directions in `experiments/parse_results.ipynb`.

Notice that all hyper-parameters in our experiments are handled by the `ml_collections` package [here](https://github.com/google/ml_collections). You can directly change those hyper-parameters at the place they are defined, e.g. `experiments/configs/individual_xxx.py`. However, a recommended way of passing different hyper-parameters -- for instance you would like to try another model -- is to do it in the launch script. Check out our launch scripts in `experiments/launch_scripts` for examples. For more information about `ml_collections`, please refer to their [repository](https://github.com/google/ml_collections).

## Reproducibility

A note for hardware: all experiments we run use one or multiple NVIDIA A100 GPUs, which have 80G memory per chip. 

We include a few examples people told us when reproducing our results. They might also include workaround for solving a similar issue in your situation. 

- [Prompting Llama-2-7B-Chat-GGML](https://github.com/llm-attacks/llm-attacks/issues/8)
- [Possible Naming Issue for Running Experiments on Windows](https://github.com/llm-attacks/llm-attacks/issues/28)

Currently the codebase only supports training with LLaMA or Pythia based models. Running the scripts with other models (with different tokenizers) will likely result in silent errors. As a tip, start by modifying [this function](https://github.com/llm-attacks/llm-attacks/blob/main/llm_attacks/base/attack_manager.py#L130) where different slices are defined for the model.

## Citation
If you find this useful in your research, please consider citing:

```
@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}
}
```

## License
`llm-attacks` is licensed under the terms of the MIT license. See LICENSE for more details.