Defect-based Testing for Safety-critical ML Components

The input space of machine learning (ML) components used in safety-critical applications is complex. Due to practical restrictions, testing such components on exponentially large datasets that cover all potential real-world situations to meaningfully measure performance metrics, such as the false negative rate, is infeasible. Hence, we must limit the test data while adequately covering critical input data points. Inspired by defect-based software testing, a methodology for specifying adequate test cases for software components, we propose a process for collecting adequate test data for ML components. Concretely, we propose systematically employing different existing ML data quality metrics and methods for enhancing the test data to uncover critical scenarios where the ML component may be less performant. We exemplify the usage of our process in two case studies involving ML components implementing different functionalities, i.e., stop sign recognition and railway track segmentation.