Node injection for class-specific network poisoning.

Area of Science:

• Graph Neural Networks
• Adversarial Attacks
• Machine Learning Security

Background:

• Graph Neural Networks (GNNs) excel at learning network representations but are vulnerable to adversarial attacks.
• Node injection attacks are practical, requiring no manipulation of existing network structures.
• Existing attacks lack specificity and stealth, making them detectable.

Purpose of the Study:

• To propose a novel class-specific node injection attack targeting GNNs.
• To develop a method (NICKI) for misclassifying specific nodes into a different class.
• To ensure injected nodes are camouflaged as benign, maintaining graph integrity.

Main Methods:

• NICKI employs an optimization-based strategy for targeted GNN sabotage.
• The attack operates in two phases: learning node representations and generating malicious node features/edges.
• Extensive experiments were conducted on four benchmark networks.

Main Results:

• NICKI significantly outperforms four baseline attacking strategies in misclassifying target nodes.
• Injected nodes are effectively camouflaged, evading detection based on topological properties.
• The poisoned graphs remain topologically indistinguishable from clean graphs.

Conclusions:

• NICKI presents an effective and stealthy class-specific adversarial attack against GNNs.
• The proposed method highlights the vulnerability of GNNs to sophisticated node injection attacks.
• Further research is needed to develop robust defenses against such sophisticated poisoning strategies.