Internet Traffic Classification Using Machine Learning Methods

This work was carried out as part of the validation of a research master's degree in computer networks and telecommunications at the University of Abomey-Calavi. The program began in February 2020 with theoretical courses, followed by a research internship starting in June.

The six-month internship took place from June to December 2020 at LETIA, under the supervision of Dr. Ratheil Houndji and Dr. Michel Dossou. Due to the COVID-19 pandemic and related restrictions, the thesis defense was postponed and eventually held on November 17, 2021, allowing additional analyses to be conducted.

The thesis is available online: master's thesis (PDF, in French).

Automatic Internet traffic classification is a key challenge for network management, security and quality of service. While many machine learning-based approaches exist, most rely on idealized datasets and overlook the presence of missing features.

The objective of this work is to design robust and efficient classification models capable of operating under partial information. A progressive feature analysis strategy was proposed to identify and eliminate redundant or weakly informative variables.

• Development of traffic classification models capable of distinguishing multiple categories of network flows while remaining robust to partial information.
• In-depth analysis of 249 features (TCP ports, transmission durations, packet sizes, effective bandwidth, etc.) describing 377,526 network flows (web, multimedia, databases, messaging, games, file transfer, attacks, P2P, remote connections) to identify the most influential attributes for classification.
• Selection of an optimal subset of 14 essential features, leading to more interpretable and less complex models.
• Achievement of classification accuracies up to 98% using supervised learning algorithms such as decision trees, random forests and XGBoost.