Design for dependability — State of the art and trends

Abstract

This paper presents an overview of design for dependability as a process involving three distinct but interrelated activities: risk analysis, risk mitigation, and risk assessment. Although these activities have been the subject of numerous works, few of them address the issue of their integration into rigorous design flows. Moreover, most existing results focus on dependability for small-size safety-critical systems with specific static architectures. They cannot be applied to large systems, such as autonomous systems with dynamic heterogeneous architectures and AI components. The overwhelming complexity and lack of interpretability of AI present challenges to model-based techniques and require empirical approaches. Furthermore, it is impossible to cope with all potential risks at design time; run-time assurance techniques are necessary to cost-effectively achieve the desired degree of dependability. The paper synthesizes the state of the art showing particularly the impact of new trends stemming from the integration of AI components in design flows. It argues that these trends will have a profound impact on design methods and the level of dependability. It advocates the need for a new theoretical basis for dependability engineering that allows the integration of traditional model-based approaches and data-driven techniques in the search for trade-offs between efficiency and dependability.