数据与计算发展前沿 ›› 2025, Vol. 7 ›› Issue (2): 3-11.

CSTR: 32002.14.jfdc.CN10-1649/TP.2025.02.001

doi: 10.11871/jfdc.issn.2096-742X.2025.02.001

• 专刊：中国科技云10周年 • 上一篇    下一篇

一种基于深度学习的轻量化流量识别模型

李勇1,2(),任勇毛1,2,殷卓然1,2,周旭1,2,*()   

  1. 1.中国科学院计算机网络信息中心，北京 100083
    2.中国科学院大学，北京 100049
  • 收稿日期:2025-02-20 出版日期:2025-04-20 发布日期:2025-04-23
  • 通讯作者: 周旭
  • 作者简介:李勇，中国科学院计算机网络信息中心，硕士研究生，主要研究方向为深度学习、计算机网络。
    本文承担工作为：模型设计，模型算法实现。
    LI Yong is a Master’s student at the Computer Network Information Center, Chinese Academy of Sciences. His main research interests include deep learning and computer networks.In this paper, he is responsible for model design and implementation of model algorithms.
    E-mail: liyong@cnic.cn|周旭，中国科学院计算机网络信息中心，博士生导师，主要研究方向为计算机网络、分布式网络架构等。
    本文承担工作为：指导优化模型和模型设计。
    ZHOU Xu, is a doctoral supervisor at the Computer Network Information Center, Chinese Academy of Sciences. His main research interests include computer networks and distributed network architectures.
    In this paper, he is responsible for guiding the optimization of the model and model design.
    E-mail: zhouxu@cstnet.cn
  • 基金资助:
    中国科学院湖南省联合攻关项目(2024JK4001);北京市自然科学基金项目(1232011)

A Lightweight Traffic Identification Model Based on Deep Learning

LI Yong1,2(),REN Yongmao1,2,YIN Zhuoran1,2,ZHOU Xu1,2,*()   

  1. 1. Computer Network Information Center, Chinese Academy of Sciences, Beijing 100083, China
    2. University of Chinese Academy of Sciences, Beijing 100049, China
  • Received:2025-02-20 Online:2025-04-20 Published:2025-04-23
  • Contact: ZHOU Xu

RichHTML

3

PDF

18

摘要：

【目的】针对深度学习流量识别模型通常面临较高的参数量和计算量，难以在资源受限的边缘网络环境中部署，提出了一种基于ShuffleNetV2改进的轻量化流量识别模型。【方法】首先，选择合适的模型分辨率因子和宽度因子来平衡模型的效率和性能；其次，在模型中嵌入轻量级多尺度特征融合模块和改进的坐标注意力机制模块，旨在以较小的计算存储开销提升模型的识别性能；最后，将模型中的激活函数ReLU替换成ReLU6，更有利于模型在资源受限的环境中部署与推理。【结果】实验基于公开的ISCXVPN2016数据集，结果验证了该模型在识别精度和执行效率上的良好表现。与现有广泛采用的ResNet-18模型相比，该模型在准确率（98.8%）增加0.2%的情况下，参数量和计算量分别大幅下降96.46%和62.97%。【结论】模型在保持较高的准确率的同时，显著实现了轻量化。

关键词: 流量识别, ShuffleNetV2, 注意力机制, 特征融合

Abstract:

[Objective] To address the issue of deep learning-based traffic identification models typically having high parameter count and computational complexity, which limits their deployment in resource-constrained edge networks, a lightweight traffic identification model based on an improved ShuffleNetV2 is proposed. [Methods] First, appropriate resolution and width factors are selected to balance the model’s efficiency and performance. Second, a lightweight multi-scale feature fusion module and an enhanced coordinate attention mechanism module are integrated into the model to improve identification performance with minimal computational and storage overhead. Finally, the ReLU activation function is replaced with ReLU6 to enhance the model’s suitability for deployment and inference in resource-limited settings. [Results] The experiments are conducted based on the publicly available ISCXVPN2016 dataset, and the results validate the model's excellent performance in terms of recognition accuracy and execution efficiency. Compared to the widely adopted ResNet-18 model, this model achieves an accuracy increase of 0.2% (reaching 98.8%), while significantly reducing the parameter count and computational complexity by 96.46% and 62.97%, respectively. [Conclusions] The model achieves significant lightweighting while maintaining high accuracy.

Key words: traffic identification, ShuffleNetV2, attention mechanism, feature fusion