Specifying Functional Safety Requirements for AI/ML in terms of Metamorphic Relations
Abstract
The specification of functional safety requirements is an integral part of a safety engineering process, especially in the case of a system whose behavior is largely determined by software. Many industry standards, such as RTCA DO-178C (aerospace), ISO 26262 (automotive) and CENELEC EN 50128 (rail), underscore the importance of specifying functional safety requirements. However, for the development of software that uses Machine Learning (ML) or other forms of Artificial Intelligence (AI), there is a widespread view that the functionality that results from the use of AI/ML cannot be specified with the level of precision that would normally be expected for the specification of functional safety requirements. A possible solution to this conundrum is specifying functional safety requirements in the form of Metamorphic Relations (MR). In this form, requirements specify a constraint on how a change to a situation or a scenario should affect the behavior of the system. Specifying functional safety requirements in the form of MRs accommodates the inherent limitations on how precisely the predictable behavior of a system can be specified. This also creates the opportunity to use Metamorphic Testing (MT) which can be an effective means of safety validation for software developed using AI/ML.