Scholarlink

Graph Convolutional Networks (GCNs) are increasingly used in security-sensitive applications. This study investigates the vulnerability of GCNs increasingly employed in IoT traffic classification, to backdoor adversarial attacks, focusing on simulated Malaria DoS and Slowite attacks. Subtle manipulations of edge features, constituting less than 0.5% perturbation, result in significant misclassifications of malicious traffic while causing less than a 1% change in overall metrics. Using a proprietary IoT dataset and validated on the CICIoMT2024 dataset, the study demonstrates consistent adversarial efficacy while preserving performance on clean data. The findings highlight the need for robust defenses to protect GCNs in critical IoT contexts, where undetected attacks could cause severe real-world impacts.

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

Backdoor Adversarial Machine Learning Attack on Graph Convolutional Networks for IoMT Traffic Misclassification

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Michael Georgiades,
  • Faisal Hussain,
  • Lakis Christodoulou

摘要

Graph Convolutional Networks (GCNs) are increasingly used in security-sensitive applications. This study investigates the vulnerability of GCNs increasingly employed in IoT traffic classification, to backdoor adversarial attacks, focusing on simulated Malaria DoS and Slowite attacks. Subtle manipulations of edge features, constituting less than 0.5% perturbation, result in significant misclassifications of malicious traffic while causing less than a 1% change in overall metrics. Using a proprietary IoT dataset and validated on the CICIoMT2024 dataset, the study demonstrates consistent adversarial efficacy while preserving performance on clean data. The findings highlight the need for robust defenses to protect GCNs in critical IoT contexts, where undetected attacks could cause severe real-world impacts.