isneslab / Wendigo

Code library for the Wendigo framework from 'Wendigo: Deep Reinforcement Learning for Denial-of-Service Query Discovery in GraphQL'

---

Wendigo

Deep Reinforcement Learning for Denial-of-Service Query Discovery in GraphQL

This repository is the official release of the code used for the Wendigo Paper published in Deep Learning Security and Privacy Workshop (DSLP) 2024, co-located with IEEE S&P.

If you plan to use this repository in your projects, please cite the following paper:

@inproceedings{mcfadden2024wendigo,
  title = {Wendigo: Deep Reinforcement Learning for Denial-of-Service Query Discovery in GraphQL},
  author = {McFadden, Shae and Maugeri, Marcello and Hicks, Chris and Mavroudis, Vasilis and Pierazzi, Fabio},
  booktitle = {Proc. of the {IEEE} Workshop on Deep Learning Security and Privacy ({DLSP})},
  year = {2024},
}

---

Disclaimer

Please note that the code in this repository is only a research prototype and may generate damaging queries. Do not use against any real systems without prior, written consent of the targets. This code is released under a "Modified (Non-Commercial) BSD License": see the terms here.

---

Use Instructions

1. python3 -m venv venv
2. source venv/bin/activate
3. pip3 install -r requirements.txt
4. pip3 install .
5. cd Wendigo
6. python3 main.py

---

Wendigo Code Breakdown

The Wendigo code consists of the following sections:

• attack: The code and results used to run the DoS impact evaluation which utilized the found query to perform a mock DoS attack.
• environments: The environment is broken down into a parent and subclass.
  • GraphQLEnv.py: handles the docker and connection,
  • GraphQLDoSEnv.py: performs handles the mappings between agent's state/action and queries.
• models: This directory contains the code for PPO and Random.
  • PPO.py: Contains the PPO code utilized for the evaluation of the paper and is a modified version of cleanrl's PPO implementation.
  • Random.py: Contains the implementation used for random evaluation of the paper (note: greedy is a setting in the attack settings).
• results: This directory contains pickles of the results presented in the paper.
• schemas: This contains the DVGA schema used to inform the environment for state to query space mappings.
• settings: This directory contains the settings files used in the evaluation (Has its own read me for clarity on the settings).
• utils: Contains helper functions and functions to interface between the agent and environment.
• main.py: This file is used to run the application by specifying the variables below the application will load the appropriate setting file and run the experiment.
  • MODEL: 'PPO' or 'Random'
  • TEST: currently only 'DoS' is supported
  • DESC: 'Regular', 'Large', 'Greedy-Regular' or 'Greedy-Large'