Enabling AI Bill of Materials through Graph-Structured Metadata: A Model-Based Systems Engineering Approach
Abstract
The use of artificial intelligence in critical cyberphysical and software-intensive systems creates new supply chain assurance challenges. Conventional Software Bills of Materials describe software components but do not adequately represent the datasets, training activities, model versions, and inference events that influence AI behavior. This paper applies Model-Based Systems Engineering to define a reference architecture for capturing and managing metadata collected across AI system lifecycle and software supply chain. The architecture combines SysML model elements with an ontology backed knowledge graph that links datasets, training runs, models and inference events. A proof-of-concept SAE J1939 anomaly detector was evaluated using CAN traffic collected from a controlled testbench comprising three physical electronic control units. The traffic was augmented with controlled cansend injections that emulated additional source-address and NAME combinations and introduced documented attack scenarios. The evaluation demonstrates the capture of lifecycle metadata and the tracing of a false negative inference event to its associated model, training run, and poisoned dataset version. The results demonstrate the feasibility of using graph structured metadata as a traceable foundation for AI Bills of Materials.