User equipment traffic classification in the 5G core

Abstract

The fifth generation of cellular network (5G), specified by the 3 rd Generation Partnership Program (3GPP), distinguishes itself from previous mobile network generations primarily through the adoption of a Service Based Architecture (SBA). By leveraging the flexibility of this new architectural paradigm, this work investigates the feasibility of implementing a Network Data Analytics Function (NWDAF) mechanism for classifying 5G devices. Following the 3GPP specifications, the International Telecommunication Union (ITU) defines three 5G service axes: Enhanced Mobile Broadband (eMBB), Ultra Reliable and Low Latency Communications (URLLC), and Massive Machine-Type Communications (mMTC). These axes focus on applications that require high throughput, low latency with high availability, and low-energy transmissions with thousands of connected Internet of Things (IoT) devices, respectively. Therefore, the classification of different device types is crucial, as each 5G axis is associated with distinct constraints, which influence resource allocation in the 5G Core (5GC). Defined by the 3GPP in Release 15 as responsible for data analysis in 5G networks, the NWDAF remains considerably underexplored in existing literature. To address this gap, this work employs a simulated 5G environment utilizing Free/Libre and Open Source Software (FLOSS). This environment incorporates UERANSIM, a Radio Access Network (RAN) and User Equipment (UE) simulator; free5GC, a 5GC implementation; and a custom network traffic generator. Two real-world 5G datasets were used to train eleven different Machine Learning (ML) models — Linear Regression (LR), Histogram-based Gradient Boosting (HGB), Light Gradient Boosting Machine (LGBM), Multilayer Perceptron (MLP), Random Forest (RF), Linear Support Vector Classification (SVC), eXtreme Gradient Boosting (XGB), Decision Tree (DT), Adaboost, Stacking, and Voting — to classify UEs in the axes based on their observed network traffic. In the implemented ML pipeline the model inference performance results achieved up to 99.91% F1-score for the eMBB class using RF, LGBM, and XGB. For the URLLC class the best results were 99.99% F1-score with SVC. In contrast, due to the sparse data points of the mMTC class, while Adaboost achieved a 99.99% F1-score in the burst scenario, the performance was limited to 2.74% in the probabilistic scenario. Despite using techniques that include balancing the training dataset, the models suffered from overfitting. Inspired by the open science movement, both the software used and the dataset created for inference are publicly accessible. This approach ensures reproducibility and establishes a foundation for future investigations into data analysis as specified by 3GPP.