Project A05: Consistency of Data-Defined Models
Many CPS now include data-defined models (e.g., ML/Deep NN components) whose evolving, opaque behavior is hard to verify and cannot be exhaustively tested. The project treats this as a consistency problem between specified behavior and actual run-time behavior, ensuring safety via a safety shield (simplex/sandbox/enforcer). The shield leverages V-SUM insights (e.g., braking force, reaction times) as assumptions, and feeds back findings, enabling run-time assurance rather than purely design-time proofs. This integrates design-time and run-time verification/validation, allowing data-defined components to be handled like classical models w.r.t. safety properties. A KA‑RaceIng case study illustrates reachable-set–based planning of fail-safe trajectories that are continuously recomputed and parameterized by V-SUM knowledge. The project will also define suitable formalisms to express safety as a consistency relation between specification and monitored data.