Scholarlink

This paper investigates the problem of defending deep graph neural networks against backdoor attacks, a critical challenge in secure AI systems. With the widespread use of pre-trained models and open-source datasets, data contamination poses a significant threat to deep learning models. To address this issue, we propose a Data-Cleaning-Based Defense algorithm (DCBD) that monitors the dynamic difference of training loss between clean data and poisoned data during GNN training to effectively remove poisoned data. Notably, DCBD identifies poisoned data at an early stage in the training process, terminating learning on poisoned inputs. Our algorithm does not require prior knowledge of the proportion of contaminated data, making it more practical for real-world applications. Experimental results show that DCBD achieves high data cleaning accuracy and outperforms baseline algorithms in both attack success rate reduction and downstream tasks’ performance.

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

Guarding Graph Neural Networks Against Backdoor Attacks-A Training Loss Dynamics Approach

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Dongqi Wang,
  • Ruiguo Hu,
  • Dongming Chen,
  • Zeyu Lv,
  • Jie Wang

摘要

This paper investigates the problem of defending deep graph neural networks against backdoor attacks, a critical challenge in secure AI systems. With the widespread use of pre-trained models and open-source datasets, data contamination poses a significant threat to deep learning models. To address this issue, we propose a Data-Cleaning-Based Defense algorithm (DCBD) that monitors the dynamic difference of training loss between clean data and poisoned data during GNN training to effectively remove poisoned data. Notably, DCBD identifies poisoned data at an early stage in the training process, terminating learning on poisoned inputs. Our algorithm does not require prior knowledge of the proportion of contaminated data, making it more practical for real-world applications. Experimental results show that DCBD achieves high data cleaning accuracy and outperforms baseline algorithms in both attack success rate reduction and downstream tasks’ performance.