Boosting Graph Robustness Against Backdoor Attacks: An Over-Similarity Perspective
Researchers have developed a new defense method called SimGuard to protect Graph Neural Networks (GNNs) from backdoor attacks. The method leverages an over-similarity observation, noting that malicious triggers in GNNs often share excessive similarity in features and structure with clean nodes. SimGuard uses a similarity-based metric to detect these triggers and employs contrastive learning to train a detector that effectively separates malicious triggers from legitimate data, thereby preserving the network's performance on clean nodes. AI
IMPACT Enhances the security and reliability of Graph Neural Networks, potentially enabling wider adoption in sensitive applications.