Scholarlink

<p>Multi-task network traffic anomaly detection could indicate the inability to model non-Euclidean traffic structures and negative transfer between tasks. To address these important challenges, we propose a novel synergistic framework. Unlike previous graph convolutional networks (GCN) relied on fixed geometric structures or computationally intensive capsule routing methods, this framework combines Curvature-aware GCN (CGCN) and an attention-based capsule network to achieve adaptive structural modeling and efficient multi-task discrimination. CGCN compensates for the lack of explicit topology in sequence data. Combined with an attention mechanism, the lightweight Capsule Network achieves collaborative fusion of multi-task features. Moreover, BERT is introduced in the framework to resolve the semantic gap in traffic. Experimental results on three public datasets could demonstrate that the proposed method achieves leading performance across various tasks, with F1-scores ranging from 98.96% to 99.92%, representing a 1–3% performance improvement compared to mainstream benchmark models. This research may provide an effective solution for implementing parallel multi-task anomaly detection in complex network environments.</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

A capsule network approach for traffic anomaly detection based on enhanced representation learning

  • 期刊论文

  • 发表日期：

    2026年3月5日

查看全文
  • Yingjing Wang,
  • Lan Li

摘要

Multi-task network traffic anomaly detection could indicate the inability to model non-Euclidean traffic structures and negative transfer between tasks. To address these important challenges, we propose a novel synergistic framework. Unlike previous graph convolutional networks (GCN) relied on fixed geometric structures or computationally intensive capsule routing methods, this framework combines Curvature-aware GCN (CGCN) and an attention-based capsule network to achieve adaptive structural modeling and efficient multi-task discrimination. CGCN compensates for the lack of explicit topology in sequence data. Combined with an attention mechanism, the lightweight Capsule Network achieves collaborative fusion of multi-task features. Moreover, BERT is introduced in the framework to resolve the semantic gap in traffic. Experimental results on three public datasets could demonstrate that the proposed method achieves leading performance across various tasks, with F1-scores ranging from 98.96% to 99.92%, representing a 1–3% performance improvement compared to mainstream benchmark models. This research may provide an effective solution for implementing parallel multi-task anomaly detection in complex network environments.