Graph Neural Backdoor: Fundamentals, Methodologies, Applications, and Future Directions

• DOI: 10.48550/arxiv.2406.10573
• Publication Date: 2024-06-15
• Authors: Xiao Yang; Gaolei Li; Jianhua Li
• Keywords: Backdoor

Graph Neural Networks (GNNs) have significantly advanced various downstream graph-relevant tasks, encompassing recommender systems, molecular structure prediction, social media analysis, etc. Despite the boosts of GNN, recent research has empirically demonstrated its potential vulnerability to backdoor attacks, wherein adversaries employ triggers poison input samples, inducing GNN adversary-premeditated malicious outputs. This is typically due controlled training process, or deployment untrusted models, such as delegating model third-party service, leveraging external sets, and employing pre-trained models from online sources. Although there's an ongoing increase in on backdoors, comprehensive investigation into this field lacking. To bridge gap, we propose first survey dedicated backdoors. We begin by outlining fundamental definition followed detailed summarization categorization current attacks defenses based their technical characteristics application scenarios. Subsequently, analysis applicability use cases backdoors undertaken. Finally, exploration directions presented. aims explore principles graph provide insights defenders, promote future security research.