Architectural Assurance of AI-Based Driving Functions in Software-Defined Vehicles

With the ever-growing size and complexity of software used in vehicles, one differentiating factor perceivable for the user is the advanced functionality offered by AI-enabled software, e.g., emergency braking or lane keeping assistants. As always in the development of vehicles, safety is a paramount requirement for such software components. Analyses of the safety of such software components often rely on assumptions about the performance of the components and the usage of AI induces uncertainty, and the high resource demand of neural networks requires new forms of hardware components and system architectures. In addition, the development process of the software becomes more agile, requiring over-the-air updates and frequent analysis of compatibility and requirement fulfillment. To overcome the challenges arising from these developments regarding the analysis of AI-enabled systems, this paper proposes a new multi-level prediction system, that applies the concepts of model-based system engineering by employing architectural analysis and multi-level simulation. To gather the challenges stemming from these developments we reviewed the state of the art in the architectural analysis of AI-enabled systems in regard to safety and performance and how such systems can be assured and derived eight different challenges from them.