GNN Robustness Toolbox (GRT)

The GNN Robustness Toolbox (GRT) is a Python framework for evaluating the robustness of Graph Neural Network (GNN) models against adversarial attacks. GRT provides a flexible and extensible platform for conducting robustness experiments, enabling researchers and practitioners to:

• Systematically evaluate the robustness of different GNN architectures.
• Compare the effectiveness of various adversarial attack strategies.
• Develop and benchmark defense mechanisms against adversarial attacks.

Key Features:

• Extensible Architecture: Easily integrate custom models, attacks, datasets, transforms, optimizers, and loss functions.
• Flexible Configuration: Define experiments using a user-friendly YAML configuration file.
• Model Caching: Cache trained models and results to avoid redundant computations.
• Comprehensive Output: Generate detailed experiment results in JSON and CSV format.

Installation:

git clone https://github.com/Munkhtenger19/Toolbox.git
cd Toolbox
pip install -r requirements.txt

Usage:

1. Define Custom Components (Optional): Create and register custom models, attacks, datasets, etc., in the custom_components directory.

2. Configure Experiments: Create a YAML configuration file specifying the experiment settings (see configs/ for examples).

3. Run Experiments: Execute the main.py script with the configuration file path:

   python main.py --cfg path/to/config.yaml

License:

GRT is released under the MIT License.