Graph-based Interpretable Anomaly Detection Framework for Network Slice Management in Beyond 5G Networks

• Published in: NOMS 2023-2023 IEEE/IFIP Network Operations and Management Symposium pp. 1 - 6
• Main Authors: Chawla, Ashima, Bosneag, Anne-Marie, Dalgkitsis, Anestis
• Format: Conference Proceeding
• Language: English
• Published: IEEE 08-05-2023
• Subjects: 5G mobile communication; Automation; Beyond 5G; eXplainable AI; Neural Network; Probabilistic logic; Recurrent neural networks; Root cause analysis; Time series analysis; Traffic control
• ISSN: 2374-9709
• DOI: 10.1109/NOMS56928.2023.10154357

5G and Beyond 5G systems are built with unprecedented scale, heterogeneity and programmability in mind, and come with features such as increased technological complexity (due to increased heterogeneity and abstraction); dynamic behavior of the network prompted by dynamic demand for new functionality and capacity extensions; variability and volume of network access (e.g., type and number of connections, diverse types of services and traffic patterns), etc. Therefore, one of the major network management requirements is automation. Rather than having only local, segregated automation mechanisms (e.g., node self-configuration, SON), it is now agreed that end-to-end proactive automation is a must. In this paper, we present a novel solution for supporting automated and proactive decision-making at the network slice level in 5G/ Beyond 5G telecommunication systems. We propose a Graph-based Interpretable Anomaly Detection (G5IAD) framework that combines the use of Graph Convolutional Neural Networks with model correlations between multivariate network slice Key Performance Indicators (KPIs) and forecasts individual time series using Recurrent Neural Networks/GRUs. The architecture also includes an anomaly detection mechanism based on probabilistic classification. This unsupervised approach helps network operators identify & explain features of the KPIs, leading to faster root cause analysis and enabling self-managed network slices. Our model was tested in a realistic 5G simulation and demonstrated 45.9% better accuracy when trained with a Graph CNN and Recurrent Neural Network compared to the GRU model.