Machine Learning–Based Mitigation of Confidentiality Violations in Software Architectures

Modern software systems are increasingly complex and interconnected, demanding rigorous analysis to ensure properties such as confidentiality. However, uncertainty in systems and environments limits precise architecture-based confidentiality analysis and hampers automated model repair. Existing approaches detect confidentiality violations but fail to mitigate them effectively. This paper introduces a machine learning–enhanced analysis that evaluates the criticality of violations and automates their repair, bridging the gap between detection and mitigation. Our evaluation shows that logistic regression best ranks uncertainty sources, and, combined with incremental testing, our approach outperforms the state of the art with up to 60× faster runtimes.