A Case Study on defining traceable Machine Learning Safety Requirements for an Automotive Perception component

Integrating machine learning (ML) into safety-critical systems introduces significant safety assurance challenges, particularly as these systems become increasingly autonomous and operate in more open and complex environments. One of the most significant of these challenges is how to systematically specify traceable ML safety requirements. In this paper, we explore the challenges of specifying safety requirements for ML components through a case study of a vehicle Automated Lane Centering function, in which an ML model performs lane detection in a highway scenario. We show how safety concerns propagate from system-level hazards, and explore specific issues that arise in defining meaningful and traceable ML-level requirements, including specifying ML behaviour and robustness. The paper provides the first detailed case study showing how effective and traceable ML safety requirements can be specified for an ML component.