A quality attribute like confidentiality is critical to trustworthy software but unfortunately, very challenging to ensure. This is because modern software systems are complex and interconnected. Architecture-based confidentiality analysis enables the early detection of violations, helping to mitigate risks before deployment. However, uncertainty in software systems and their environments complicates precise and comprehensive architectural analysis. Additionally, the complexity of software models and the exponential growth of uncertainty scenarios pose significant challenges for automated mitigation, often leaving software architects to resolve confidentiality violations manually, a process that is both time-intensive and error-prone.
In this paper, we extend our machine-learning-based approach to mitigate confidentiality violations. Specifically, we introduce a novel mitigation strategy inspired by TCP Congestion Control, as well as a strategy that capitalizes on clustering techniques to dynamically adjust batch sizes. Our evaluation on three real-world software architectures demonstrates that our extended approach can mitigate confidentiality violations while outperforming the state-of-the-art. Whereas previously the upper limit was 60 times runtime reduction, now we achieve 2298 times reduction, with the median being an elevenfold reduction. Our statistical analysis confirms that the added TCP-inspired strategy is significantly cheaper than the state-of-the-art baseline (Friedman test and Nemenyi post hoc test ), while also having a strong practical impact (Kendall’s W ). This extended work deepens our understanding of the nature of uncertainty and also of the techniques optimally suited to mitigating the violations caused by uncertainties. It takes us one step closer to designing trustworthier systems.
Zur Publikation