Integrating Explainable AI to Enhance Dynamic Risk Assessment in Automated Driving Systems

This paper introduces a Dynamic Risk Assessment (DRA) framework for Automated Driving Systems (ADS) that integrates Explainable Artificial Intelligence (XAI) to enhance the transparency and interpretability of Machine Learning (ML) models. Traditional risk assessment methods, such as those outlined in ISO 26262 and ISO 21448, typically focus on the concept phase and rely on static risk models. However, these approaches are inadequate for addressing the evolving, runtime risks faced by ADS during operation. To overcome these limitations, we propose incorporating runtime risk assessment using the PDTAA (Plan-Do-Train-Adjust-Assess) process, which enables continuous risk monitoring as driving conditions change. Through a case study on Adaptive Cruise Control (ACC) systems in a Highway Lane-Following scenario, we demonstrate how the framework allows for detailed, continuous risk evaluation. The integration of XAI techniques, specifically Shapley values, improves the interpretability and reliability of risk predictions, offering a more robust solution for runtime safety evaluation in ADS. This work contributes to the application of XAI to DRA, offering a valuable tool to enhance the safety and reliability of ADS before deployment in real-world environments.