Constructing adversarial examples to investigate the plausibility of explanations in deep audio and image classifiers

Abstract

Given the rise of deep learning and its inherent black-box nature, the desire to interpret these systems and explain their behaviour became increasingly more prominent. The main idea of so-called explainers is to identify which features of particular samples have the most influence on a classifier’s prediction, and present them as explanations. Evaluating explainers, however, is difficult, due to reasons such as a lack of ground truth. In this work, we construct adversarial examples to check the plausibility of explanations, perturbing input deliberately to change a classifier’s prediction. This allows us to investigate whether explainers are able to detect these perturbed regions as the parts of an input that strongly influence a particular classification. Our results from the audio and image domain suggest that the investigated explainers often fail to identify the input regions most relevant for a prediction; hence, it remains questionable whether explanations are useful or potentially misleading.