Network Traffic Classification using Deep Neural Networks

Abstract

Network traffic control engineers are actively engaged in the exploration of methods to analyze network traffic for the purpose of detecting protocols, services, and applications. This task holds significant importance for Internet Service Providers (ISPs) and Telecommunication Companies (Telcos) as they need to monitor the types of traffic protocols traversing their networks in order to shape their Quality of Service (QoS) strategies and marketing approaches. Prior research efforts have identified machine learning techniques that offer potential in identifying various applications and services within network traffic. Building upon this foundation, a novel research endeavor has been undertaken, involving the creation of a meticulously curated dataset. This dataset was meticulously assembled by capturing network flows on a university network across different time intervals. The dataset was meticulously designed to encompass a comprehensive range of network flows and protocols. The dataset subsequently underwent a rigorous pre-processing phase, involving the extraction of feature values, removal of incomplete flows, and the application of Server Name Indication (SNI) based labeling. The study focused on three core high-level features, namely packet characteristics, payload attributes, and inter-arrival times. To derive insights from this dataset, an array of sophisticated deep learning architectures grounded in Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) were employed. To provide a comprehensive benchmark, a traditional Random Forest Classifier was also integrated into the study for comparative purposes. The model training process was executed using the proprietary dataset, yielding exceptional accuracy levels of up to 99%. Notably, a thorough evaluation was conducted, directly comparing the performance of the deep learning model against the Random Forest approach. The outcomes clearly demonstrated the superiority of the deep learning (DL) model over the conventional Machine Learning (ML) approach, exemplifying the potential of advanced DL techniques in network traffic analysis and protocol detection. This research contributes to the growing body of knowledge in network traffic management, offering insights that could shape future strategies in traffic analysis and control.