Jailbreaking Starter: Adversarial Robustness for Language Models

This project implements an adversarial testing framework for evaluating the robustness of large language models (LLMs). It utilizes methods like Monte Carlo Tree Search (MCTS) with Double Progressive Widening (DPW) and adversarial suffix generation to test and optimize prompts against predefined ethical and moderation constraints.

Features

• Adversarial Testing: Evaluate LLM responses against adversarially crafted prompts.
• Monte Carlo Tree Search (MCTS): Includes a DPW solver for efficient exploration of adversarial space.
• Support for Multiple Models: Works with models such as GPT, Vicuna, and LLaMa.
• Moderation API: Integrates OpenAI's Moderation API for content filtering.
• Visualization: Includes tools for loss tracking and visualization.

Code Structure

• KOV.py: The main script managing the adversarial testing process and framework setup.
• WhiteBox.py: Implements the white-box optimization methods, including token sampling and adversarial suffix generation.
• DPW_Solver.py: Contains the DPW-based MCTS solver for optimizing adversarial search.
• utils.py: Utility functions for model loading, prompt generation, moderation checks, and loss visualization.

Requirements

• Python 3.8 or higher
• PyTorch
• Transformers
• OpenAI API (requires OPENAI_API_KEY environment variable)
• Additional dependencies listed in the requirements.txt file.

Installation

1. Clone the repository:

   git clone [https://github.com/your-username/jailbreaking-starter.git](https://github.com/phantom2810/Jailbreak-Simulator-Adversarial-Testing-for-LLMs.git)

2. Install dependencies:

   pip install -r requirements.txt

3. Set up your OpenAI API key:

   export OPENAI_API_KEY=your-api-key

Usage

Running the Framework

1. Configure the input parameters and models in KOV.py.
2. Run the main script:

   python KOV.py

Example Output

The framework evaluates adversarial prompts and generates scores for LLM responses:

• Moderation category scores
• Adversarial success rates
• Loss and perplexity metrics

Project Files

• KOV.py: Main script for orchestrating the adversarial testing process.
• WhiteBox.py: Implements the white-box Markov Decision Process (MDP) for adversarial prompt generation.
• DPW_Solver.py: Provides MCTS logic with progressive widening for state-action exploration.
• utils.py: Helper functions for prompt processing, scoring, and response generation.

Contributing

Contributions are welcome! If you encounter issues or have ideas for improvement, feel free to open an issue or submit a pull request.