PyGIP 0.1.0 documentation

PyGIP is a comprehensive Python library focused on model extraction attacks and defenses in Graph Neural Networks (GNNs). Built on PyTorch, PyTorch Geometric, and DGL, the library offers a robust framework for understanding, implementing, and defending against attacks targeting GNN models.

PyGIP is featured for:

Extensive Attack Implementations: Multiple strategies for GNN model extraction attacks, including fidelity and accuracy evaluation.
Defensive Techniques: Tools for creating robust defense mechanisms, such as watermarking graphs and inserting synthetic nodes.
Unified API: Intuitive APIs for both attacks and defenses.
Integration with PyTorch/DGL: Seamlessly integrates with PyTorch Geometric and DGL for scalable graph processing.
Customizable: Supports user-defined attack and defense configurations.

Quick Start Example:

Outlier Detection Example with 5 Lines of Code:

from gnn_mae import ModelExtractionAttack2

dataset = ...  # Load your graph dataset as a DGL object
attack = ModelExtractionAttack2(dataset, attack_node_fraction=0.25)
attack.attack()

# Evaluate fidelity and accuracy
print(f"Fidelity: {attack.fidelity}, Accuracy: {attack.accuracy}")

Attack Modules¶

Class Name	Attack Type	Reference
AdvME	Model Extraction Attack; Black-Box	DeFazio, D., & Ramesh, A. (2019). Adversarial model extraction on graph neural networks. arXiv preprint arXiv:1912.07721.
GNNMEA0	Model Extraction Attack; Grey-Box	Wu, B., Yang, X., Pan, S., & Yuan, X. (2022). Model extraction attacks on graph neural networks: Taxonomy and realisation. In Proceedings of the 2022 ACM on Asia conference on computer and communications security, 337-350.
GNNMEA1	Model Extraction Attack; Black-Box	Wu, B., Yang, X., Pan, S., & Yuan, X. (2022). Model extraction attacks on graph neural networks: Taxonomy and realisation. In Proceedings of the 2022 ACM on Asia conference on computer and communications security, 337-350.
GNNMEA2	Model Extraction Attack; Black-Box	Wu, B., Yang, X., Pan, S., & Yuan, X. (2022). Model extraction attacks on graph neural networks: Taxonomy and realisation. In Proceedings of the 2022 ACM on Asia conference on computer and communications security, 337-350.
GNNMEA3	Model Extraction Attack; Black-Box	Wu, B., Yang, X., Pan, S., & Yuan, X. (2022). Model extraction attacks on graph neural networks: Taxonomy and realisation. In Proceedings of the 2022 ACM on Asia conference on computer and communications security, 337-350.
GNNMEA4	Model Extraction Attack; Grey-Box	Wu, B., Yang, X., Pan, S., & Yuan, X. (2022). Model extraction attacks on graph neural networks: Taxonomy and realisation. In Proceedings of the 2022 ACM on Asia conference on computer and communications security, 337-350.
GNNMEA5	Model Extraction Attack; Grey-Box	Wu, B., Yang, X., Pan, S., & Yuan, X. (2022). Model extraction attacks on graph neural networks: Taxonomy and realisation. In Proceedings of the 2022 ACM on Asia conference on computer and communications security, 337-350.
GNNStealing	Model Extraction Attack; Black-Box	Shen, Y., He, X., Han, Y., & Zhang, Y. (2022). Model stealing attacks against inductive graph neural networks. In 2022 IEEE Symposium on Security and Privacy (SP), 1175-1192.
*Grey-box attacks assume the attacker has at least partial knowledge across all dimensions (i.e., node attributes, graph structure, and shadow graph).
**Black-box attacks refer to scenarios where the attacker has no knowledge regarding at least one dimension (i.e., node attributes, graph structure, or shadow graph).

Defense Modules¶

Class Name	Defense Type	Reference
watermarkMED	Watermarking-based defense technology. Defend against model extraction attack	Zhao, X., Wu, H., & Zhang, X. (2021). Watermarking graph neural networks by random graphs. In 2021 9th International Symposium on Digital Forensics and Security (ISDFS), 1-6.

Additional Details¶

For further information, refer to individual classes like GraphNeuralNetworkMetric, WatermarkGraph, or the specific model extraction attacks in the source code.