Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

On the Robustness of Graph Reduction Against GNN Backdoor

Yuxuan Zhu, Michael Mandulak, Kerui Wu, George Slota, Yuseok Jeon, Ka-Ho Chow, Lei Yu

TL;DR

The paper investigates how graph reduction techniques influence GNN backdoor robustness, revealing that coarsening can mitigate backdoor ASR while sparsification can inadvertently elevate vulnerability. Using a broad empirical study across datasets, architectures, and reduction methods, the authors show that ASR reductions from coarsening are most pronounced on larger graphs and for GraphSAGE/GAT, with VC/VE often yielding the best mitigation. They provide detailed trigger and poisoned-node analyses to explain why reductions disrupt backdoor signals, particularly for low-degree nodes, and demonstrate that blindly applying sparsification may compromise security despite memory gains. The findings emphasize integrating robustness considerations into graph-reduction pipelines to balance scalability with security in real-world GNN deployments.

Abstract

Graph Neural Networks (GNNs) are gaining popularity across various domains due to their effectiveness in learning graph-structured data. Nevertheless, they have been shown to be susceptible to backdoor poisoning attacks, which pose serious threats to real-world applications. Meanwhile, graph reduction techniques, including coarsening and sparsification, which have long been employed to improve the scalability of large graph computational tasks, have recently emerged as effective methods for accelerating GNN training on large-scale graphs. However, the current development and deployment of graph reduction techniques for large graphs overlook the potential risks of data poisoning attacks against GNNs. It is not yet clear how graph reduction interacts with existing backdoor attacks. This paper conducts a thorough examination of the robustness of graph reduction methods in scalable GNN training in the presence of state-of-the-art backdoor attacks. We performed a comprehensive robustness analysis across six coarsening methods and six sparsification methods for graph reduction, under three GNN backdoor attacks against three GNN architectures. Our findings indicate that the effectiveness of graph reduction methods in mitigating attack success rates varies significantly, with some methods even exacerbating the attacks. Through detailed analyses of triggers and poisoned nodes, we interpret our findings and enhance our understanding of how graph reduction influences robustness against backdoor attacks. These results highlight the critical need for incorporating robustness considerations in graph reduction for GNN training, ensuring that enhancements in computational efficiency do not compromise the security of GNN systems.

On the Robustness of Graph Reduction Against GNN Backdoor

TL;DR

The paper investigates how graph reduction techniques influence GNN backdoor robustness, revealing that coarsening can mitigate backdoor ASR while sparsification can inadvertently elevate vulnerability. Using a broad empirical study across datasets, architectures, and reduction methods, the authors show that ASR reductions from coarsening are most pronounced on larger graphs and for GraphSAGE/GAT, with VC/VE often yielding the best mitigation. They provide detailed trigger and poisoned-node analyses to explain why reductions disrupt backdoor signals, particularly for low-degree nodes, and demonstrate that blindly applying sparsification may compromise security despite memory gains. The findings emphasize integrating robustness considerations into graph-reduction pipelines to balance scalability with security in real-world GNN deployments.

Abstract

Graph Neural Networks (GNNs) are gaining popularity across various domains due to their effectiveness in learning graph-structured data. Nevertheless, they have been shown to be susceptible to backdoor poisoning attacks, which pose serious threats to real-world applications. Meanwhile, graph reduction techniques, including coarsening and sparsification, which have long been employed to improve the scalability of large graph computational tasks, have recently emerged as effective methods for accelerating GNN training on large-scale graphs. However, the current development and deployment of graph reduction techniques for large graphs overlook the potential risks of data poisoning attacks against GNNs. It is not yet clear how graph reduction interacts with existing backdoor attacks. This paper conducts a thorough examination of the robustness of graph reduction methods in scalable GNN training in the presence of state-of-the-art backdoor attacks. We performed a comprehensive robustness analysis across six coarsening methods and six sparsification methods for graph reduction, under three GNN backdoor attacks against three GNN architectures. Our findings indicate that the effectiveness of graph reduction methods in mitigating attack success rates varies significantly, with some methods even exacerbating the attacks. Through detailed analyses of triggers and poisoned nodes, we interpret our findings and enhance our understanding of how graph reduction influences robustness against backdoor attacks. These results highlight the critical need for incorporating robustness considerations in graph reduction for GNN training, ensuring that enhancements in computational efficiency do not compromise the security of GNN systems.
Paper Structure (28 sections, 6 figures, 18 tables)

This paper contains 28 sections, 6 figures, 18 tables.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Graph neural network (GNN)
  4. Graph Reduction
  5. Graph Poisoning and Backdoor Attack
  6. Methodology
  7. System Assumption
  8. Attack Model
  9. Evalution Framework
  10. Experiment
  11. Experimental Settings
  12. GNN Training with Graph Reduction
  13. Robustness Effect Analysis for Coarsening
  14. Attack mitigation with graph coarsening
  15. ASR and ACC v.s. Coarsening ratio
  16. ...and 13 more sections

Figures (6)

  • Figure 1: Backdoor attack on GNN under graph reduction.
  • Figure 2: Impact of Graph Coarsening on ASR with Trigger Size 3 and Poisoning Ratio 5%.
  • Figure 3: Impact of Graph Coarsening on ASR with different GNN models. A light-colored line represents the baseline ASR under no coarsening.
  • Figure 4: ASR Under Different Graph Coarsening Methods.
  • Figure 5: ASR Under Different Graph Sparsification Methods.
  • ...and 1 more figures