ARTEMIS - Adaptive Response and Tracking of External Model Inputs in Safety-Critical Systems

In Cyber-Physical System (CPS), neural networks are increasingly employed to process high-dimensional, heterogeneous sensor data and support real-time, automated decisionmaking. These systems, ranging from autonomous vehicles to smart grids and medical devices, rely heavily on the reliability and accuracy of neural models. However, one of the central challenges in deploying neural networks in real-world CPS is their vulnerability to Out of Distribution (OOD) data-input samples that differ significantly from the training data. OOD data can arise due to unforeseen environmental changes, sensor malfunctions, adversarial conditions, or simply evolving operating contexts. When neural networks encounter such data, their predictions often become unreliable, which can lead to severe consequences, especially in safety-critical applications. This paper presents a structured process for integrating OOD monitoring into CPS model pipelines, addressing the full lifecycle of detection, analysis, and adaptation. A central focus of the paper is the structured handling of detected OOD data. This allows to assess which inputs are most critical to system performance and safety. ... mehrFurthermore, the paper explores the theoretical underpinnings of incremental model updates using OOD data, while maintaining model stability and compliance with realtime constraints. Although the proposal is theoretical, it lays the groundwork for a practical bridge between OOD detection research and its integration into CPS. Emphasis is placed on monitoring pipelines that trace specific samples throughout the system lifecycle, offering insights into long-term reliability. Ultimately, the paper advocates for a proactive and adaptive approach to handling OOD data in alignment with the system’s Operational Design Domain (ODD), positioning this integration as a critical factor for ensuring the long-term safety, reliability, and adaptability of CPS.