Failure On Demand Analysis in the Case of Score Based Binary Classifiers: Method and Application

Safety assessment and verification have become more complex in the past years. Especially the incorporation of machine learning components, and their black box nature, are proposing new difficulties to overcome. Therefore new techniques are needed to judge the safety of machine learning components and further integrate those into existing safety analysis methods. In this contribution we will provide a new method for safety analysis of a score based binary classifier. The presented technique can output a single reliable value for the failure on demand. Latter one can then be used inside a system safety analysis, as done for physical engineering systems. In particular we will briefly mention a general approach for score based binary classifiers, as already applied for general systems. Furthermore we will contribute a more refined method in the case of a normal distributed score. The main idea is to incorporate confidential bounds on the parameters to obtain a function that serves as upper bound for the failure on demand. Further analysis of the retrieved function will then provide a mathematically based single value for the reliability. In the end of this work we will demonstrate this technique at the example of breast cancer detection and evaluate the performance in this scenario.

Keywords: Safety analysis, Failure, Binary classifier, Normal distribution, Score, Confident bounds.

Download PDF