PhD Dissertation Defense by: Inoussa Mouiche

The School of Computer Science is pleased to present…

Date: Monday, November 24, 2025

Time: 10am – 1pm

Abstract: Modern cyber defense relies on timely and accurate Cyber Threat Intelligence (CTI), yet essential information about threat actors, malware, vulnerabilities, and attack patterns remains buried in unstructured reports. This dissertation presents a unified suite of machine-learning frameworks that transform raw CTI into structured, machine-readable knowledge to enable automated reasoning, decision support, real-time detection, and large-scale analysis. First, we introduce a data-centric framework that automates the integration of heterogeneous CTI datasets using the STIX standard together with normalization, alias expansion, and fuzzy matching. This produces scalable, harmonized training corpora essential for reliable model development. Second, we advance pipeline-based CTI extraction by integrating security-aware contextual embeddings, novel neural architectures, and ontology-guided error control, substantially reducing noise propagation and improving entity and relation extraction accuracy. Third, we propose a joint extraction model based on a Multisequence Labeling Representation (MSLR) enriched with expert domain features, enabling robust handling of overlapping relations, feature confusion, and language ambiguity—achieving state-of-the-art results on the DNRTI-JE benchmark. Finally, we develop a data-centric relation extraction framework that leverages simplified architectures to outperform more complex model-centric approaches. Collectively, these contributions deliver reliable, explainable, and scalable knowledge modeling solutions, paving the way for real-time, AI-driven cybersecurity intelligence with broad applicability across safety-critical domains.