A dynamic ensemble learning model for robust Graph Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph Neural Networks (GNNs) face significant vulnerabilities from adversarial attacks.
• Existing defense models struggle with dynamic data and attack variations, limiting real-world applicability.

Purpose of the Study:

• To develop a robust defense mechanism for GNNs against adversarial attacks in dynamic environments.
• Introduce DERG, a dynamic ensemble learning model for enhanced GNN security.

Main Methods:

• Graph sampling strategy to purify perturbed graphs and generate diverse subgraphs.
• Mutual information-based diversity enhancement to increase submodel variability.
• Game theory-based decision strategy for dynamic submodel weighting and adaptation.

Main Results:

• DERG demonstrates significant robustness against diverse adversarial attacks.
• Effective defense against graph modification, backdoor poisoning, and double attacks.
• Adaptability to changing environments through dynamic submodel selection.

Conclusions:

• DERG offers a novel and effective solution for defending GNNs against adversarial threats.
• The dynamic ensemble approach enhances GNN resilience in unpredictable scenarios.
• Future work could explore further optimizations and applications of DERG.