GraphGuard: An adaptive approach for restoring accuracy in backdoor-compromised GNNs

Area of Science:

• Artificial Intelligence
• Machine Learning Security

Background:

• Backdoor attacks threaten machine learning models, especially Graph Neural Networks (GNNs).
• Existing defenses often focus on detection rather than accurate restoration of model performance.
• The complex structure of graph data complicates GNN defense strategies.

Purpose of the Study:

• To develop a method for restoring the original accuracy of GNNs compromised by backdoor attacks.
• To enhance GNN resilience against hidden triggers and poisoned inputs.
• To improve the interpretability of GNN decision-making post-attack.

Main Methods:

• Combining advanced filtering to remove suspicious data points and augmentation to strengthen GNNs against triggers.
• Implementing an adaptive framework to balance filtering and augmentation based on attack severity and model sensitivity.
• Integrating Explainable AI (XAI) techniques for transparent detection and understanding of backdoor triggers.

Main Results:

• Achieved an average accuracy restoration of 97-99% across various backdoor attack scenarios.
• Demonstrated effective reduction of false positives and negatives in backdoor detection.
• Enhanced GNN integrity and performance in the presence of sophisticated attacks.

Conclusions:

• The proposed method offers an effective solution for restoring GNN accuracy after backdoor attacks.
• The adaptive filtering and augmentation strategy significantly improves model resilience.
• XAI integration enhances transparency and trust in GNN security.