Safety Assessment: From Black-Box to White-Box

Abstract

Safety assurance for Machine-Learning (ML) based applications such as object detection is a challenging task due to the black-box nature of many ML methods and the associated uncertainties of its output. To increase evidence in the safe behavior of such ML algorithms an explainable and/or interpretable introspective model can help to investigate the black-box prediction quality. For safety assessment this explainable model should be of reduced complexity and humanly comprehensible, so that any decision regarding safety can be traced back to known and comprehensible factors. We present an approach to create an explainable, introspective model (i.e., white-box) for a deep neural network (i.e., black-box) to determine how safety-relevant input features influence the prediction performance, in particular, for confidence and Bounding Box (BBox) regression. For this, Random Forest (RF) models are trained to predict a YOLOv5 object detector output, for specifically selected safety-relevant input features from the open context environment. The RF predicts the YOLOv5 output reliability for three safety related target variables, namely: softmax score, BBox center shift and BBox size shift. The results indicate that the RF prediction for softmax score are only reliable within certain constrains, while the RF prediction for BBox center/size shift are only reliable for small offsets.