DTGBA: A stronger graph backdoor attack with dual triggers.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Cybersecurity

Background:

• Graph neural networks (GNNs) are vulnerable to backdoor attacks.
• Existing attacks lack imperceptibility and are susceptible to defenses like random edge dropping.

Purpose of the Study:

• To propose a novel graph backdoor attack, Dual Triggers Graph Backdoor Attack (DTGBA).
• To enhance the imperceptibility and robustness of graph backdoor attacks against existing defenses.

Main Methods:

• DTGBA utilizes an imperceptible injected trigger generator with discriminators for adversarial training.
• A feature mask learner identifies and modifies key feature dimensions to create feature-based triggers.
• Dual triggers (injected and feature-based) ensure attack effectiveness even after trigger removal.

Main Results:

• DTGBA demonstrates superior performance in degrading GNN accuracy.
• The proposed attack is more imperceptible and robust compared to existing methods.
• Extensive experiments validate the effectiveness of DTGBA.

Conclusions:

• DTGBA presents a significant advancement in graph backdoor attack methodologies.
• The dual-trigger approach overcomes limitations of previous attacks and defenses.
• This research highlights critical security vulnerabilities in GNNs.