Evaluating the Validity of Word-level Adversarial Attacks with Large Language Models

AVLLM

This guide provides detailed steps for fully fine-tuning the AVLLM model using the LLaMA Factory framework.
Follow these instructions to set up your environment, prepare your model and datasets, and perform fine-tuning and evaluations.

Environment Setup

1. Clone the LLaMA Factory Repository

   Clone the repository to your home directory:

   git clone https://github.com/hiyouga/LLaMA-Factory.git

2. Create and Activate a Conda Environment

   Create a new Conda environment named AVLLM and activate it:

   conda create -n AVLLM python=3.10  
   conda activate AVLLM  

3. Install Dependencies

   Navigate to the LLaMA-Factory directory and install the required dependencies:

   cd LLaMA-Factory  
   pip install -r requirements.txt

Model and Data Preparation

1. Download the Model

   Download the TinyLlama model from Hugging Face and place it in the model_path directory:

   • Model URL: TinyLlama-1.1B-intermediate-step-1431k-3T

2. Prepare the Datasets

   Download your fine-tuning dataset AVLLM_train.json and testing dataset AVLLM_test.json. Load these datasets into the LLaMA-Factory/data directory and update the dataset_info.json file to include:

   "AVLLM-train": {
   "file_name": "AVLLM_train.json"
   },
   "AVLLM-test": {
   "file_name": "AVLLM_test.json"
   }

Fine-Tuning

1. Configure the Fine-Tuning Script

   In the llama_factory directory, create or modify the full-ft.sh script.

2. Configure DeepSpeed
   Create or modify the deepspeed.json file.

3. Execute the Fine-Tuning Script
   Run the fine-tuning script:

   sh full-ft.sh  

Adversarial Attack Module Setup

1. Install TextAttack
   Install TextAttack for generating adversarial examples:

   pip install textattack

2. Generate Adversarial Examples
   Use the following command to generate adversarial examples for evaluation:

   sh generate_example.sh

3. Prepare Evaluation Data

   -Store the generated adversarial examples as attack_example.csv.
   -Process this file using data_processing.py to format it for AVLLM inference.
   -Load the resulting evaluation.json into the LLaMA-Factory/data directory and update dataset_info.json:

   "AVLLM-evaluation": {
   "file_name": "evaluation.json"
   }

4. Run Inference and Evaluation

   Use the predict.sh script for inference, and then evaluate the ASR with evaluation.py.

   sh predict.sh  
   python evaluation.py  

Patch Module Setup

1. Add Custom Module
   Place your module in textattack/constraints/semantics/tinyllama.py and update textattack/attack_args.py to include your module in the CONSTRAINT_CLASS_NAMES:

   "tinyllama": "textattack.constraints.semantics.Tinyllama"

2. API Call Setup
   Set up an API call by running server.py and adjust the address in tinyllama.py accordingly.

   python server.py

3. Run Attack Command
   Execute the following command to run your attack module:

   sh patch.sh  

📝 Citation

If you find this paper useful, please consider staring 🌟 this repo and citing 📑 our paper:

@inproceedings{zhou2024evaluating,
  title={Evaluating the validity of word-level adversarial attacks with large language models},
  author={Zhou, Huichi and Wang, Zhaoyang and Wang, Hongtao and Chen, Dongping and Mu, Wenhan and Zhang, Fangyuan},
  booktitle={Findings of the Association for Computational Linguistics ACL 2024},
  pages={4902--4922},
  year={2024}
}