The convergence of heterogeneous wireless access technologies, particularly Wi-Fi and 5G, introduces new security challenges at the network access level. These challenges require Intrusion Detection Systems (IDSs) to operate under strict performance, resource, and deployment constraints. Although Artificial Intelligence (AI)-based IDSs increasingly dominate the literature, their effectiveness and practical deployability in real-world wireless environments, particularly Wi-Fi networks, remain limited. These limitations stem from shortcomings in performance evaluation, feature representation, and architectural design.
This thesis studies AI-based intrusion detection mechanisms in wireless networks, with a particular focus on Wi-Fi as a critical enabler of future Wi-Fi/5G convergence. A comparative analysis of state-of-the-art detection methods and architectures highlights the prevalence of learning-based approaches while identifying key gaps that hinder their adoption in operational settings.
To address these gaps, this work first proposes a multi-criteria evaluation framework that captures complementary dimensions of detection performance, including security effectiveness, computational complexity, and uncertainty. Building on this framework, a Pareto-based tradeoff selection method is applied to guide the selection of optimal configurations across security, efficiency, and reliability dimensions.
Next, a hybrid statistical and information-theoretic feature selection approach, based on mutual information and a weighted Chi²-test of independence, is introduced to derive a compact, IEEE 802.11–aware feature set for efficient detection of MAC-layer attacks. Experimental results demonstrate that well-selected features significantly reduce model complexity and enable shallow detection models to outperform deeper architectures under equivalent conditions.
Finally, the selected feature set is leveraged within a federated learning–based intrusion detection architecture at the network edge. The proposed solution achieves strong generalization and high detection accuracy across heterogeneous devices while preserving privacy, supporting adaptability, and maintaining computational feasibility.
Collectively, the contributions of this thesis provide an end-to-end, practical approach to AI-enabled intrusion detection in industrial Wi-Fi networks and establish a foundation for secure, scalable, and interoperable intrusion detection in future converged Wi-Fi/5G environments.