The Impact of Feature Quality on SDN Traffic Learning and Classification

Network traffic classification is crucial for facilitating sophisticated network management. Software-defined networks (SDN) provide a featured architecture where the control and data planes are decoupled to enable dynamic configuration. The control plane functionality is implemented in a programmable centralized controller that enables effective use of data-driven based on network traffic classification using efficient machine learning (ML) algorithms. However, there is a lack of standardized benchmarking criteria agreed by the SDN community to evaluate the techniques proposed to collect and classify the network traffic features. This paper presents an empirical analysis to investigate the impact of SDN traffic features on the performance of ML-based classifiers. This study has been conducted using a publicly available SDN traffic dataset. The procedure of feature engineering has also been comprehensively evaluated through three approaches of feature selection, which are low variance filter, high correlation filter, and backward feature elimination. The experimental results show that the performance of ML classifiers is highly driven by the quality traffic features fed to them; the highest quality the traffic features are the best accuracy is achieved.