Scholarlink

<p>In response to the challenges of weak adaptability to real-world environments and deployment difficulties of current deep learning methods for network traffic anomaly detection, this research investigates the features of network traffic. The study uses Long Short-Term Memory networks and one-dimensional Convolutional Neural Networks to capture the spatial and temporal correlations of data, enabling the model to learn sequence dependencies and complex features. This enhances the model’s learning capability and generalization performance, while improving its robustness to noise and changes in data distribution. The paper also proposes a heterogeneous knowledge distillation model combined with Generative Adversarial Networks, as well as an isomorphic cyclic distillation architecture. Through self-supervision and self-adjustment mechanisms, these approaches reduce the model size and improve detection accuracy. The proposed model has been validated through both machine learning and deep learning comparisons, showing its practical value. By evaluating the model on the dataset, the reduction in Floating Point Operations and Parameters falls within the range of 97 to 99%.</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

Anomaly traffic detection in heterogeneous lightweight networks based on spatio-temporal features

  • 期刊论文

  • 发表日期：

    2025年4月2日

查看全文
  • Qingyun Liang,
  • Ligang Cong,
  • Heng Du

摘要

In response to the challenges of weak adaptability to real-world environments and deployment difficulties of current deep learning methods for network traffic anomaly detection, this research investigates the features of network traffic. The study uses Long Short-Term Memory networks and one-dimensional Convolutional Neural Networks to capture the spatial and temporal correlations of data, enabling the model to learn sequence dependencies and complex features. This enhances the model’s learning capability and generalization performance, while improving its robustness to noise and changes in data distribution. The paper also proposes a heterogeneous knowledge distillation model combined with Generative Adversarial Networks, as well as an isomorphic cyclic distillation architecture. Through self-supervision and self-adjustment mechanisms, these approaches reduce the model size and improve detection accuracy. The proposed model has been validated through both machine learning and deep learning comparisons, showing its practical value. By evaluating the model on the dataset, the reduction in Floating Point Operations and Parameters falls within the range of 97 to 99%.