Novel Explanation-Based Method Detects Backdoor Attacks in Graph Neural Networks

Researchers have developed a new detection method to identify backdoor attacks in Graph Neural Networks (GNNs), addressing critical security vulnerabilities in AI systems. While GNNs are widely used, they remain susceptible to malicious intrusions that compromise performance and ethical standards. Existing detection techniques often rely on rigid, single-metric approaches that fail to capture the complexity of backdoor behaviors. This study introduces an innovative approach leveraging graph-level explanations by extracting and transforming secondary outputs from GNN explanation mechanisms. The team devised seven novel metrics to effectively detect these attacks and created an adaptive attack model to rigorously evaluate their method. Tested across multiple benchmark datasets against various attack models, the proposed solution demonstrated high detection performance. This advancement significantly enhances the reliability and security of GNN classification tasks, offering a robust safeguard against sophisticated backdoor threats in machine learning applications.