Implementing packet-level internet traffic classification using XGBoost in software-defined networking

The continuous growth of Internet applications with varying degrees of Quality of Service (QoS) requirements poses significant challenges to network administrators and Internet service providers in network management and QoS provisioning. Software-defined networking (SDN) aims to overcome the limitations of traditional network management by separating network control from data forwarding. Traffic classification (TC) using machine learning (ML) has attracted the attention of researchers due to its ability to provide valuable insights into ongoing network flows. While SDN offers programmable centralized control, implementing packet-based TC within the control plane introduces substantial processing overhead on SDN controllers. This paper presents an offloading TC framework that uses ML to classify bidirectional network traffic, preserving controller efficiency while providing insights into network behavior. Using the Unicauca-dataset-2019, comprehensive preprocessing and data engineering methods were applied, focusing on early-stage TC. The proposed eXtreme Gradient Boosting (XGBoost) model achieved 91.77% classification accuracy across 11 unique network traffic classes, representing traffic patterns of 37 Internet applications, using lightweight statistical features in early-stage packet inspection. The offloading approach moves the standard TC workflow to the data plane, reducing the controller’s burden of packet-level feature extraction and classification. Experimental testbed evaluation showed that the offloaded TC approach significantly outperformed control plane TC, offered performance comparable to simple switching, and reduced the number of ingress packets to the controller by 72%.