On the Properness of Incorporating Binary Classification Machine Learning Algorithms Into Safety-Critical Systems

Manufacturers are willing to incorporate Machine Learning (ML) algorithms into their systems, especially those considered as Safety-Critical Systems (SCS). ML algorithms that perform binary classification (i.e., Binary Classifiers (BCs)) find a wide applicability as error, intrusion or failure detectors, provided that their performance complies with SCS safety requirements. However, the performance analysis of BCs relies on metrics that were not developed with safety in mind and consequently may not provide meaningful evidence to decide whether to incorporate a BC into a SCS. In this paper, we empirically assess the properness of such incorporation by analyzing the distribution of misclassifications of BCs instead of simply counting misclassifications. This allows us to better assess the adequacy of a given BC by identifying areas of the classification space where the BC is likely to misclassify and therefore constitutes actionable information to deal with the SCS. Our assessment takes a deeper view of the classification performance concerning safety by using new metrics that consider the proportions of predictions that are/are not considered sufficiently safe to be used by incorporating SCS. The results of our experiment allow discussing the potential of such distribution analysis for deciding if a BC can be incorporated into a SCS.