On the Necessity of Explicit Artifact Links in Safety Assurance Cases for Machine Learning

Abstract

The perception in autonomous systems is essential for safe behavior. Machine learning (ML)-based functions play an increasingly important role in this context. The development and safety assurance of such functions is different from the development of non-ML-based functions. Traceability of the various artifacts generated for safety argumentation is challenging, as there is i.e. no longer a direct mapping from requirements to code and data cannot be directly mapped to a semantic domain model. In this work, we show that and how the links between artifacts, which are created in different stages of the development, must be established explicitly. These links enable us to build confidence in our safety argumentation. We concretize these explicit links in two examples, namely pedestrian detection and vehicle detection.