An Investigation of Machine Learning Solutions to Demonstrate Black Box Software Conformity in Regulated Measuring Instruments

Modern regulated measuring instruments are mostly driven by sophisticated software with simple hardware sensors, where the software conformity with legal requirements is usually demonstrated via the version numbers and cryptographic hashes over executable binaries. Such conformity assessments quickly reach a bottleneck, when the actual software functionalities remain hidden from the view of authorities, i.e., in black box scenarios. In this article, we develop extensions to a machine learning approach that allows software modeling as automata in quasi-black box settings. Via evaluation of performance and comparison with other state-of-the-art approaches, we further propose a novel machine learning solution that includes the strengths of automata learning and supervisedly trained artificial neural networks, offering simpler conformity tests and enabling potential general functional identifications of regulated software-driven measuring instruments from both behavioral and metrological perspectives. By providing requirements and preliminaries, it is envisioned that this article could build a theoretical foundation in legally regulated sectors and beyond, such that manufacturers or authorities can monitor practical software-driven devices for abnormalities without performing new conformity assessments.