Add surrogate extraction v0.2 for GNN models with class-based implementation

[Shah621]

Surrogate Extraction Fidelity Results for GNN Models

This document presents the fidelity results of surrogate models when varying the number of attack nodes for three popular datasets: Cora, Citeseer, and PubMed.

Results Summary

Dataset: Cora

Number of Attack Nodes	Fidelity (%)
140	85.38%
280	85.23%
420	83.57%
560	83.16%
700	83.12%

Dataset: Citeseer

Number of Attack Nodes	Fidelity (%)
140	78.21%
280	80.13%
420	78.48%
560	78.36%
700	79.35%

Dataset: PubMed

Number of Attack Nodes	Fidelity (%)
140	88.22%
280	88.13%
420	88.91%
560	89.21%
700	88.71%

Analysis

• Cora shows a slight decline in fidelity as the number of attack nodes increases.
• Citeseer exhibits variability, with fidelity peaking at 280 attack nodes.
• PubMed maintains consistently high fidelity, with a slight increase as the number of attack nodes increases to 560.

These results suggest that the PubMed dataset is more robust to surrogate extraction attacks compared to Cora and Citeseer.

Conclusion

The fidelity of surrogate models depends on the dataset characteristics and the number of attack nodes. The results can help in understanding the robustness of Graph Neural Networks (GNNs) to surrogate extraction attacks.