Cyber Resilience and Privacy in Healthcare Information Systems Using Machine Learning

Healthcare software systems occupy a uniquely sensitive intersection of patient safety, personal data protection, and regulatory compliance, making the security of their code a matter of technical and ethical urgency. Although privacy engineering and automated vulnerability detection have progressed over the years, both have mostly developed separately. Privacy frameworks address concerns at the architectural and design level, while vulnerability detection tools focus on code, lacking a systematic way to connect the two. This separation creates gaps in secure software development, as privacy threats found during system design cannot be linked to specific weaknesses in code. Additionally, vulnerability detection tools are not tailored or tested for healthcare software, meaning development teams lack integrated, domain-specific tools that their sensitive systems require. This thesis addresses that gap through five interconnected contributions. First, the LINDDUN privacy threat modelling methodology is systematically extended by constructing a mapping between all seven LINDDUN threat categories (Linkability, Identifiability, Non-repudiation, Detectability, Disclosure of Information, Unawareness, and Non-compliance) and their corresponding Common Weakness Enumeration (CWE) entries and thus operationalising the translation from design-time privacy threats to code-level weaknesses. The mapping was validated through structured expert assessment, achieving substantial inter-rater agreement (Cohen's κ = 0.78), establishing it as a reliable analytical instrument. Second, a large-scale, healthcare-specific vulnerability dataset is constructed comprising 6,308,314 annotated C/C++ functions drawn from DiverseVul, ReposVul, and StarCoder, spanning 773 distinct CWE identifiers and enriched with LINDDUN privacy threat labels. This makes it the first dataset to integrate privacy threat taxonomy with software vulnerability classification at this scale and domain specificity. Third, a comprehensive comparative evaluation of 14 machine learning models across three paradigms (traditional ML, graph neural networks (GNNs), and transformer-based models) was conducted on this dataset. Transformer models achieve the highest detection performance, with CodeGPT-small recording an F1 score of 0.9976 and a false positive rate of 0.09%, and DistilBERT achieving the highest recall of 0.9985. Among GNN approaches, Reveal attains an F1 of 0.9958 with an FPR of 0.15%, while Devign yields the highest GNN recall at 0.9977. Among traditional methods, XGBoost delivers an F1 of 0.9926 with an FPR of 0.40%, demonstrating that computationally efficient classical models remain highly competitive and represent a viable alternative where GPU infrastructure is unavailable. Multi-class CWE classification proves substantially more challenging across all paradigms, with Random Forest achieving the best macro-F1 of 0.676, confirming that fine-grained vulnerability typing remains an open problem. Fourth, a novel hybrid detection architecture is developed, combining deterministic template matching for known vulnerability patterns with a machine learning fallback for novel pattern detection and safety rule overrides that provide hard false-negative guarantees unattainable by purely statistical approaches. Fifth, this architecture is realised as a validated, deployment-feasible prototype. Deployed as a real time REST API with CPU-only inference feasibility, the system achieves an overall accuracy of 99.11% and delivers CWE-specific remediation guidance alongside each detection decision and is validated through user acceptance testing with industry practitioners and academic researchers. This work presents the first end-to-end, traceable framework that connects privacy threat identification during design with automated vulnerability detection in code, tailored specifically for healthcare. By using healthcare-specific CWE profiles and LINDDUN threat categories, it offers targeted tools for healthcare software teams that support security needs in regulated environments. Additionally, the results show that traditional ML and hybrid models can perform competitively with transformers, suggesting resource-efficient alternatives for effective vulnerability detection in healthcare and similar contexts.