Exploring Explainable Artificial Intelligence (XAI) in Telecommunication Networks

Abstract

Artificial intelligence (AI) is poised to become a key technology in future telecommunication networks due to its profound ability to solve complex tasks in a data-driven and, in some cases, model-free manner. Although AI has proven its ability to yield superior performance over traditional optimisation approaches in many other engineering disciplines, such as audio, video and image processing, etc., its adoption into key critical areas of the telecommunication domain remains severely hampered. One of the main barriers to the widespread adoption of AI in the telecommunication domain stems from an inherent lack of understanding of how complex AI models formulate their decisions and the potential implications this has towards the safe and reliable operation of the network. Recently, the field of eXplainable AI (XAI) has emerged as a promising area of research that can help overcome the "black-box" limitation associated with complex AI models by providing practitioners and other stakeholders with explanations as to how the AI model arrives at its final decisions. In this thesis, we explore some of the potential benefits that XAI can have within the telecommunication domain. Specifically, we focus on two key areas of the networking ecosystem, where the general complexity of the system motivates both the adoption of AI-driven solutions and where the critical nature of the problem means that transparency becomes a crucial requirement to ensure any decisions of the proposed solution can be thoroughly validated and trusted by its affected stakeholders. The main research question we address in this thesis can be summarized as follows: "How can explainable artificial intelligence (XAI) be integrated into high-stake areas of the telecommunication domain to ensure better reliability and robust service within the network?". In addressing this research question, the first part of this thesis explores the problem of network slicing, where a network operator may lease specific slices of its network resources and infrastructure to secondary tenants. Within this particular research domain, we first propose and compare various AI-based solutions for optimising the network resources scheduled to the slice tenant. We then augment our proposed framework with additional layers of explainability and demonstrate how such explanations can be used to gain important insights into the model's decisions, as well as to aid the human-in-the-loop to debug and devise targeted solutions to fix any potential misbehaviour of the model. In the second part of this thesis, we explore some of the benefits that XAI can bring within the cybersecurity space. We begin by proposing a novel methodology for explainable intrusion detection. In contrast to existing works on this topic, our proposed methodology includes an additional layer of unsupervised machine learning (ML), which takes the explanations from an initial network intrusion detection system (NIDS) as input and leverages the information gained from these explanations to automatically enhance the performance of the overall intrusion system against zero-day attacks. Finally, we extend our methodology to support general improvements in the overall accuracy of the system against any attack type and perform an in-depth analysis to evaluate how well our methodology generalizes to cases where i) multiple intrusion datasets are considered, ii) the choice of model used by the initial NIDS is varied, iii) the choice of XAI methodology is varied, and iv) we assess the ability of our methodology to generalise to unseen data distributions.