Hardening of Artificial Neural Networks for Use in Safety-Critical Applications -- A Mapping Study

Context: Across different domains, Artificial Neural Networks (ANNs) are used more and more in safety-critical applications in which erroneous outputs of such ANN can have catastrophic consequences. However, the development of such neural networks is still immature and good engineering practices are missing. With that, ANNs are in the same position as software was several decades ago. Today, standards for functional safety, such as ISO 26262 in the automotive domain, require the application of a collection of proven engineering principles and methods in the creation of software to increase its quality and reduce failure rates to an acceptable level. Objective: In the future, such a set of proven engineering methods needs to be established for the development of Artificial Neural Networks to allow their use in safety-critical applications. Method: This work takes a step in this direction by conducting a mapping study to extract challenges faced in the development of ANNs for safety-critical applications and to identify methods that have been used for the hardening of ANNs in such settings. Results: We extracted ten different challenges found to be repeatedly reported in the literature regarding the use of ANNs in critical contexts. All of these challenges are addressed by engineering methods, of which we identified 54 in our study that can be used for the hardening of networks. Conclusions: Various methods have been proposed to overcome the specific challenges of using ANNs in safety-critical applications. On the path towards defining best practices, we envision that future software engineering will need to focus on further investigating these methods and increasing the maturity and understanding of existing approaches, with the goal to develop clear guidance for proper engineering of high-quality ANNs.