Scholarlink

<p>In this work, we present a solution to address the challenges faced by Network Intrusion Detection Systems (NIDS) in handling limited labeled data and insufficient model generalization. We propose an enhanced framework called TMIDS (Tri-Model Intrusion Detection System), which integrates deep learning with semi-supervised learning techniques. The framework employs a Transformer with multi-head self-attention mechanisms and LSTM to achieve feature extraction and classification, respectively. Simultaneously, we design an improved pseudo-label generation method that leverages confidence fusion and a voting mechanism to enhance model performance, reduce dependency on labeled data, and simplify traditional feature engineering. Experimental results demonstrate that TMIDS significantly improves the learning ability on unlabeled samples and the classification accuracy compared to existing methods. On two distinct datasets, TMIDS achieved an accuracy exceeding 93% when trained on the complete labeled dataset. Even when trained with only 50% or 75% of the labeled data, the framework maintained an accuracy above 83%. Therefore, the proposed TMIDS framework exhibits stability across different datasets and different proportions of labeled data, enhancing the reliability and practicality of NIDS while providing an effective solution for addressing complex cybersecurity threats.</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

TMIDS: A tri-model based semi-supervised learning intrusion detection framework

  • 期刊论文

  • 发表日期：

    2026年6月17日

查看全文
  • Hongyu Tao,
  • Jue Chen,
  • Lizhi Wang,
  • Yirui Rao,
  • Shanshan Chen

摘要

In this work, we present a solution to address the challenges faced by Network Intrusion Detection Systems (NIDS) in handling limited labeled data and insufficient model generalization. We propose an enhanced framework called TMIDS (Tri-Model Intrusion Detection System), which integrates deep learning with semi-supervised learning techniques. The framework employs a Transformer with multi-head self-attention mechanisms and LSTM to achieve feature extraction and classification, respectively. Simultaneously, we design an improved pseudo-label generation method that leverages confidence fusion and a voting mechanism to enhance model performance, reduce dependency on labeled data, and simplify traditional feature engineering. Experimental results demonstrate that TMIDS significantly improves the learning ability on unlabeled samples and the classification accuracy compared to existing methods. On two distinct datasets, TMIDS achieved an accuracy exceeding 93% when trained on the complete labeled dataset. Even when trained with only 50% or 75% of the labeled data, the framework maintained an accuracy above 83%. Therefore, the proposed TMIDS framework exhibits stability across different datasets and different proportions of labeled data, enhancing the reliability and practicality of NIDS while providing an effective solution for addressing complex cybersecurity threats.