摘要
The increasing diversity of network terminals not only brings the exponential growth of network traffic, but also increases the complexity of user activities analysis. Packet encryption and dynamic port techniques also bring new challenges to traffic classification. Machine Learning (ML) techniques have been deployed at the control plane of programmable networks with better accuracy and expensive communication overhead. This paper proposes a traffic classification framework with ML classifier offloaded to the programmable data plane to decrease communication overhead. Considering the limitation of memory space in programmable switches, we design a storage algorithm combining hash tables and Count-Min sketch to store traffic features to be used in the ML classifier. We first use K-Means to classify flows into different categories, then we use decision tree to obtain the activity labels. Experimental results show that the proposed traffic classification framework can achieve high classification accuracy (above 93.7%) with minor performance degradation.
学科与关键词
- Distributed Multi-Task In-Network Classification on Programmable Switches by Ensemble ModelsLi, Qing; Lin, Jiaye; Xie, Guorui; Guan, Zhongxu; Luan, Zeyu; Qi, Zhuyun; Jiang, Yong*; Yuan, ZhenhuiIEEE TRANSACTIONS ON NETWORKING, 2025, 33(6): 3302-3317.
- In-Network Machine Learning Using Programmable Network Devices: A SurveyZheng, Changgang*; Hong, Xinpeng; Ding, Damu; Vargaftik, Shay; Ben-Itzhak, Yaniv; Zilberman, NoaIEEE Communications Surveys & Tutorials, 2024, 26(2): 1171-1200.被引 11
- Procedures, Criteria, and Machine Learning Techniques for Network Traffic Classification: A SurveySheikh, Muhammad Sameer; Peng, Yinqiao*IEEE Access, 2022, 10: 61135-61158.被引 17
