Safety-by-Design for Deep Learning Methods at the Example of Autonomous Driving

Kurzfassung

The development of autonomous vehicles and autonomous driving functions has increased rapidly in recent years. One driving factor for this is the recent breakthroughs in machine learning and deep learning. One problem with these methods is that they are black-box systems. If we have a system that we cannot understand precisely, this raises the question of how we can use it in safety-critical applications, such as the traffic sector. In general, safety concerns are present in every engineer while developing software. However, often, they focus too much on fulfilling the requirements of their stakeholders, so safety-relevant aspects take a back seat. In contrast, we want to consider them early in the design and concept phase. This approach is known as safety-by-design (ISO 26262). However, ensuring safety does not mean the absence of risk factors. It means that risks are considered and assessed, and a strategy exists for how best to deal with them (ISO 21448). Recently, much effort has been put into applying these results to AI components (ISO 8800). The field that deals with applying engineering safety arguments to AI models is known as AI engineering. Based on these findings, we want to model a trade-off between safety aspects and functionality and apply this to deep learning.