Scholarlink

<p>For the resource constraints of Internet of Things (IoT) devices, existing intrusion detection solutions often struggle to be effectively deployed on these platforms. To address this issue, this paper introduces a lightweight intrusion detection model for IoT, named CP-LSKD, leveraging knowledge distillation techniques. We employ Contrastive Principal Component Analysis (CPCA) during the data preprocessing phase, which enables effective dimensionality reduction of high-dimensional features. Additionally, we design a lightweight intrusion detection student model incorporating inverted residual networks and depthwise separable convolutions, ensuring efficient feature extraction while compressing the model. This paper proposes an Ladder-Structured Knowledge Distillation (LSKD). This method establishes shortcut connections between the teacher and student models to optimize the distillation process from both the knowledge transfer and label dimensions. To address class imbalance in the dataset, this article proposes an adaptive adjustment factor strategy applied to the focus loss function, which effectively improves the detection sensitivity of minority class samples by increasing attention to difficult to classify samples. After extensive experimental verification, CP-LSKD model not only has better key performance indicators than the current advanced intrusion detection model, but also achieves about 80% reduction in computational cost and parameters compared with the original model.</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 knowledge distillation-based lightweight intrusion detection method for the Internet of Things

  • 期刊论文

  • 发表日期：

    2025年10月9日

查看全文
  • Xin Yang,
  • Zhendong Wang,
  • Yang Shuxin,
  • Li Dahai,
  • Sammy Chan

摘要

For the resource constraints of Internet of Things (IoT) devices, existing intrusion detection solutions often struggle to be effectively deployed on these platforms. To address this issue, this paper introduces a lightweight intrusion detection model for IoT, named CP-LSKD, leveraging knowledge distillation techniques. We employ Contrastive Principal Component Analysis (CPCA) during the data preprocessing phase, which enables effective dimensionality reduction of high-dimensional features. Additionally, we design a lightweight intrusion detection student model incorporating inverted residual networks and depthwise separable convolutions, ensuring efficient feature extraction while compressing the model. This paper proposes an Ladder-Structured Knowledge Distillation (LSKD). This method establishes shortcut connections between the teacher and student models to optimize the distillation process from both the knowledge transfer and label dimensions. To address class imbalance in the dataset, this article proposes an adaptive adjustment factor strategy applied to the focus loss function, which effectively improves the detection sensitivity of minority class samples by increasing attention to difficult to classify samples. After extensive experimental verification, CP-LSKD model not only has better key performance indicators than the current advanced intrusion detection model, but also achieves about 80% reduction in computational cost and parameters compared with the original model.