Assuring Runtime Quality Requirements for AI-Based Components

As Artificial Intelligence makes astonishing progress, various AI-embedded applications are being built to unleash their potential. However, all technologies come with their inherent limitations in dealing with unanticipated situations, making it difficult to assure the satisfaction of critical qualities at runtime. This is partly due to the challenge of specifying requirements for quality-critical AI-based components. We argue that for a deployed AI model whose accuracy cannot be validated at runtime, an accuracy-centric specification method is not good enough to support AI application engineering practice. To address this fundamental issue, requirements engineering techniques can help, especially an NFRs-based approach has been proposed for mitigating the impacts of two types of errors caused by uncertainties, so that critical qualities can be assured in the specification of AI-based components. We have implemented our strategy by a combined use of requirements analysis techniques, including modelling goals as in goal-oriented RE and modelling of environment and problems as in problem-oriented RE. We have showcased its application on a Facial Recognition Payment (FRP) system. This work could help create a runtime engineering shield for AI-based components and move forward its application in quality essential scenarios.