On the Effectiveness of Packet Sampling for Early Stage Traffic Identification

Recent years, increasing studies have focused on finding effective machine learning models to identify traffics with the packets at the early stage. Capturing and processing the full packet sequence in a high speed network are with high computational and storage expenses. Therefore, packet sampling is usually applied to reduce the burden. In this paper, we try to evaluate the effectiveness of packet sampling for early stage traffic identification. Two real network traffic data sets and three typical classifiers are used for the study. We firstly compare the systematic and the stratified random sampling policies. Then we study the relationship between the sampling probability and the accuracy of the simple random sampling policy. The empirical study results suggest that packet sampling is effective for early stage traffic identification.