Path signature-based XAI-enabled network time series classification

Abstract

Classifying network time series (NTS) is crucial for automating network administration and ensuring cyberspace security. It enables the detection of anomalies, the identification of network attacks, and the monitoring of performance issues, thereby providing valuable support for network protection and optimization. However, modern communication networks pose challenges for NTS classification methods. These include handling large-scale and complex NTS data, extracting features from intricate datasets, and addressing explainability requirements. These challenges are particularly pronounced for complex 5G networks. Notably, explainability has become crucial for the widespread deployment of network automation for 5G networks and beyond. To tackle these challenges, we propose a path-signature-based NTS classification model called recurrent signature (RecurSig). This innovative model is designed to overcome the time-consuming feature selection problem by utilizing deep-learning (DL) techniques. Additionally, it provides a solution for addressing the black-box issue associated with DL models in network automation systems (NAS) by incorporating an explainable classification approach. Extensive experimentation on various public datasets demonstrates that RecurSig outperforms existing models in accuracy and explainability. The results indicate its potential for application in cyberspace security and automated network management, offering an explainable solution for network protection and optimization.