Abstract

This article presents a study with feasibility and performance analysis of machine learning (ML) techniques using supervised techniques for anomaly detection problems in a 5G communication network. The proposed ML models (Multilayer Perceptron, Decision Tree, and Support Vector Machine) were used to classify data into anomaly or non-anomaly based on two 5G Open Radio Access Network (O-RAN) datasets with various key performance indicators (KPIs). Furthermore, we propose a strategy that devotes to labeling anomalous situations, leveraging the t-Distributed Stochastic Neighbor Embedding (tSNE) technique atop datasets enclosing multiple KPIs. The results were significant, with an accuracy above 90% for all use cases considered.

IntroductionTo support new 5G cellular network requirements (e.g., data rates exceeding 10 Gbps, network latency under 1 ms, capacity expansion by a factor of 1,000, and energy efficiency gains), vendors have begun investigating new radio access network (RAN) architectures [17, 13, 25, 5, 23].

Open Radio Access Network (O-RAN), suggested by the O-RAN Alliance [15], stands as a promising radio technology that has gained worldwide acceptance. O-RAN is a worldwide community of operators, manufacturers,technology that has gained worldwide acceptance. O-RAN is a worldwide community of operators, manufacturers, and academic institutes [18, 1]. The vision is to rewrite the RAN industry towards establishing an open, adaptable, and intelligent RAN [15]. Artificial intelligence (AI) in machine learning (ML) will play a crucial role in the 5G networkwith particular emphasis on the O-RAN. For example, ML use can drive more efficient enhancements in 5G network planning, automation of network operations (e.g., provisioning, optimization, and fault prediction), network slicing, service quality prediction, and other applications and services [8, 3, 14, 20].

Conclusion

In this work, we present results and analyses of three ML/AI supervised approaches applied to anomaly detection: Multilayer Perceptron, Decision Tree, and Support Vector Machine. The tests were conducted on an emulation testbed concerning a network environment dataset. The unsupervised ML/AI strategy, based on t-Distributed Stochastic Neighbor Embedding (tSNE), was used to create data labels. Results associated with the accuracy of the ML/AI algorithms were obtained, suggesting an excellent performance with an accuracy of above 90% for all cases.