Adversarial Robustness of RAG with text only.

This project evaluates the adversarial robustness of RAG systems, which use LLMs for open-domain question answering task, by using NSGA-II as a multi-object optimization algorithm.

Notice: All results and demo video can be found here

Usage

Navigate to the project directory

• Install required packages: Run the following command to make the setup script executable:

  chmod +x vast_ai.sh
  ./vast_ai.sh

• Login to Hugging Face: Use the following command to authenticate and access the LLMs and DPR models:

  huggingface-cli login: <access_token>

• Run the attack: Execute the following command to run the adversarial attack:

  python main.py --reader_name llama-7b -n_iter 100 -pct_words_to_swap 0.2 --algorithm NSGAII

Repo structure

• algorithm.py: Implements the NSGA-II algorithm for multi-objective optimization.
• evaluate.py: Evaluates the adversarial robustness of RAG systems.
• fitness.py: Defines fitness functions for optimization.
• population.py: Manages the population for the genetic algorithm.
• reader.py: LLMs proccess.
• retrieval.py: Implements retrieval methods for RAG systems.
• typo_transformation.py: Applies typo transformations to generate adversarial examples.
• utils.py: Contains utility functions, including visualization tools.
• visualize.ipynb: Jupyter Notebook for visualizing results.

Inspired From

This project is inspired by multiple works and repositories:

• AttackText: Provides tools and techniques for generating adversarial examples in NLP tasks.
• pymoo: A Python library for multi-objective optimization, which inspired the implementation of NSGA-II in this project.
• Typos that Broke the RAG's Back: Genetic Attack on RAG Pipeline by Simulating Documents in the Wild via Low-level Perturbations: This paper inspired the typo transformation techniques implemented in typo_transformation.py and their idea.

Each of these sources has contributed to shaping the methodology, implementation, and experimental setup of this project.