This thesis addresses the issue of decision opacity in federated learning-based intrusion detection systems, models that achieve high accuracy at the expense of interpretability. The work focuses on the downstream analysis of the classifier, adopting a federated Network Intrusion Detection System (NIDS) for IoT networks, evaluated on the CICIoT2023 dataset, as the baseline configuration. To render the global model's predictions inspectable, the Explainable AI (XAI) methods SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) are integrated, quantifying the agreement and reproducibility of their attributions against a random baseline. The extracted explanations subsequently feed two locally deployed Large Language Models (LLMs) of equal size, one general-purpose and one specialized in the cybersecurity domain, which are compared in translating the evidence of each alert into a structured incident report.
Questa tesi affronta il problema dell'opacità decisionale nei sistemi di rilevamento delle intrusioni basati su apprendimento federato, modelli che raggiungono un'elevata accuratezza a scapito dell'interpretabilità. Il lavoro si concentra sull'analisi a valle del classificatore, assumendo come configurazione di riferimento un Network Intrusion Detection System (NIDS) federato per reti IoT valutato sul dataset CICIoT2023. Per rendere ispezionabili le predizioni del modello globale vengono integrati i metodi di Explainable AI (XAI) SHapley Additive exPlanations (SHAP) e Local Interpretable Model-agnostic Explanations (LIME), di cui si quantificano accordo e riproducibilità delle attribuzioni rispetto a una baseline casuale. Le spiegazioni estratte alimentano poi due Large Language Model (LLM) di pari dimensionalità eseguiti in ambiente locale, uno generalista e uno specializzato nel dominio della sicurezza informatica, posti a confronto nel tradurre le evidenze di ciascun allarme in un rapporto di incidente strutturato.
UN APPROCCIO SPIEGABILE AL RILEVAMENTO DEGLI ATTACCHI NELLE RETI IOT
GRANIERI, ANTONIO
2025/2026
Abstract
This thesis addresses the issue of decision opacity in federated learning-based intrusion detection systems, models that achieve high accuracy at the expense of interpretability. The work focuses on the downstream analysis of the classifier, adopting a federated Network Intrusion Detection System (NIDS) for IoT networks, evaluated on the CICIoT2023 dataset, as the baseline configuration. To render the global model's predictions inspectable, the Explainable AI (XAI) methods SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) are integrated, quantifying the agreement and reproducibility of their attributions against a random baseline. The extracted explanations subsequently feed two locally deployed Large Language Models (LLMs) of equal size, one general-purpose and one specialized in the cybersecurity domain, which are compared in translating the evidence of each alert into a structured incident report.I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14253/110199