Nascita, Alfredo (2024) Explaining and Improving DL Models for Network Traffic Analysis: Unveiling the Black Box via XAI. [Tesi di dottorato]

	Documento PDF Alfredo_Nascita_37_COMPLETO.pdf Visibile a [TBR] Amministratori dell'archivio Download (9MB) | Richiedi una copia
	Documento PDF Alfredo_Nascita_37_PARZIALE.pdf Visibile a [TBR] Amministratori dell'archivio Download (5MB) | Richiedi una copia

Abstract

The analysis of Internet traffic today presents new and complex challenges that make traditional approaches increasingly ineffective. Deep Learning (DL) approaches represent the most promising strategy, as they can handle the dynamic and heterogeneous nature of traffic. However, such approaches have limitations in transparency due to their "black-box" nature, which hinders adoption in real-world network environments. For these reasons, Explainable AI (XAI) has been introduced, aiming to shed light on DL models and enhance understanding of their decisions. This thesis focuses on the explainability of state-of-the-art approaches based on DL for Internet traffic analysis in various application contexts, such as traffic classification and cyber-attacks and anomaly detection. Specifically, several explainability aspects are analyzed to understand how DL models function and to guide their improvement from different perspectives. Initially, the role of inputs on model performance is studied, and a strategy is proposed to refine model complexity by selecting only the most important inputs. Subsequently, the reliability of the outputs produced by traffic classifiers is analyzed and improved. The analysis then shifts to locate knowledge in the core of models tied to specific concepts, such as class membership, enabling targeted knowledge manipulations and performance adjustments without additional fine-tuning. Finally, traffic classifiers trained with incremental approaches are analyzed to develop guidelines for training extensible traffic classifiers. The methodologies proposed aim to improve the integration of DL models in real-world contexts through in-depth explainability analyses. The solutions presented support a deeper understanding and improvement of network tools, making them more transparent, reliable, and suited for complex, evolving scenarios.

Downloads

Actions (login required)

Modifica documento