Overview

This repository implements and recreates the attacks described in the paper "Unveiling the Secrets without Data: Can Graph Neural Networks Be Exploited through Data-Free Model Extraction Attacks?", presented at the 33rd USENIX Security Symposium (2024). The implementation demonstrates data-free model extraction attacks on Graph Neural Networks (GNNs) without access to actual graph data or node features.

Architecture

The project is structured as follows:

/stealgnn (root directory)
    main.py
    /models
        __init__.py
        victim.py
        generator.py
        surrogate.py
    /attacks
        __init__.py
        attack1.py
        attack2.py
        attack3.py
    requirements.txt

Key Components

main.py

The entry point for running experiments. It handles dataset loading, model initialization, attack execution, and result reporting.

models/

• victim.py: Implements the victim GNN models for different datasets (Cora, Computers, Pubmed, OGB-Arxiv).

• generator.py: Contains the GraphGenerator class for creating synthetic graphs.

• surrogate.py: Implements the surrogate model that attempts to mimic the victim model.

attacks/

• attack1.py: Implements the Type I attack.

• attack2.py: Implements the Type II attack.

• attack3.py: Implements the Type III attack.

Attack Types

• Type I Attack: Uses gradients from both surrogate and estimated victim model.

• Type II Attack: Uses gradients only from the surrogate model.

• Type III Attack: Uses two surrogate models to capture more complex knowledge.

Datasets

The implementation supports four datasets:

• Cora

• Computers

• Pubmed

• OGB-Arxiv

How to Run

1. Install dependencies:

       pip install -r requirements.txt
   

2. Run an attack:

       python main.py <attack_type> <dataset_name>
   

   Where <attack_type> is 1, 2, or 3, and <dataset_name> is 'cora', 'computers', 'pubmed', or 'ogb-arxiv'.

Evaluation Metrics

• Accuracy

• Fidelity

• F1 Score

• Confusion Matrix

Output

The script generates:

• Printed statistics

• Confusion matrix plot

• Loss plot

• PDF report with detailed results

License

This project is licensed under the MIT License. See the LICENSE file in the repository for the full license text.

References

Zhuang, Y., Shi, C., Zhang, M., Chen, J., Lyu, L., Zhou, P., & Sun, L. (2024). Unveiling the Secrets without Data: Can Graph Neural Networks Be Exploited through Data-Free Model Extraction Attacks? USENIX Security Symposium, 2024. https://www.usenix.org/conference/usenixsecurity24/presentation/zhuang