PRAETORIAN: New Defense Against GNN Backdoor Attacks
Researchers propose PRAETORIAN, a defense against backdoor attacks on Graph Neural Networks (GNNs). Unlike prior methods that inspect specific subgraph patterns or node features, PRAETORIAN targets intrinsic requirements of effective backdoors. It detects attacks by analyzing internal correlations within potential trigger subgraphs and quantifying external node influence, reducing average attack success rate to 0.55%.
Key facts
- GNNs are vulnerable to backdoor attacks.
- Prior defenses can be circumvented by adaptive attackers.
- PRAETORIAN targets intrinsic requirements of effective backdoors.
- Attackers inject many trigger nodes or rely on highly influential ones.
- PRAETORIAN analyzes internal correlations within potential trigger subgraphs.
- It quantifies external node influence to identify disproportionate impact.
- Average attack success rate reduced to 0.55%.
- arXiv paper number: 2605.08278.
Entities
Institutions
- arXiv