Scholarlink

<p>Graph Neural Networks (GNNs) have shown exceptional performance in processing graph-structured data, making them indispensable for a wide range of real-world applications. However, GNNs are highly vulnerable to backdoor attacks, where adversaries inject stealthy subgraphs as triggers into a small subset of the training data. These poisoned datasets lead to compromised models that behave maliciously in the presence of these triggers while maintaining normal behavior otherwise. Detecting and removing these triggers is an immense challenge due to their subtle and subgraph-based nature, which allows them to evade traditional defenses. Graph spectral analysis has been widely applied to mitigate adversarial attacks on GNNs by leveraging spectral properties of graphs. However, less work has been performed using spectral analysis to defend against backdoor attacks on GNNs. As a result, in this paper, we propose Graph Spectral Purification (GSP), a defense mechanism based on low-pass filtering in the spectral domain. By suppressing high-frequency perturbations likely tied to backdoor triggers, GSP reduces the attack success rate while preserving the essential low-frequency structure of the graphs. Extensive experiments across five real-world datasets demonstrate that GSP effectively mitigates backdoor attacks, reducing the attack success rate by up to 100%, while maintaining the clean model accuracy of the GNN. These results suggest that leveraging spectral properties through low-pass filtering offers a practical and effective defense against backdoor attacks on GNNs.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Graph spectral purification for backdoor defence in graph neural networks

  • 期刊论文

  • 发表日期：

    2025年7月8日

查看全文
  • Shuiqiao Yang,
  • Bao Gia Doan,
  • Paul Montague,
  • Olivier De Vel,
  • Tamas Abraham,
  • Alsharif Abuadbba,
  • Ehsan Abbasnejad,
  • Seyit Camtepe,
  • Damith C. Ranasinghe,
  • Salil S. Kanhere

摘要

Graph Neural Networks (GNNs) have shown exceptional performance in processing graph-structured data, making them indispensable for a wide range of real-world applications. However, GNNs are highly vulnerable to backdoor attacks, where adversaries inject stealthy subgraphs as triggers into a small subset of the training data. These poisoned datasets lead to compromised models that behave maliciously in the presence of these triggers while maintaining normal behavior otherwise. Detecting and removing these triggers is an immense challenge due to their subtle and subgraph-based nature, which allows them to evade traditional defenses. Graph spectral analysis has been widely applied to mitigate adversarial attacks on GNNs by leveraging spectral properties of graphs. However, less work has been performed using spectral analysis to defend against backdoor attacks on GNNs. As a result, in this paper, we propose Graph Spectral Purification (GSP), a defense mechanism based on low-pass filtering in the spectral domain. By suppressing high-frequency perturbations likely tied to backdoor triggers, GSP reduces the attack success rate while preserving the essential low-frequency structure of the graphs. Extensive experiments across five real-world datasets demonstrate that GSP effectively mitigates backdoor attacks, reducing the attack success rate by up to 100%, while maintaining the clean model accuracy of the GNN. These results suggest that leveraging spectral properties through low-pass filtering offers a practical and effective defense against backdoor attacks on GNNs.