Expound: A Black-Box Approach for Generating Diversity-Driven Adversarial Examples

Deep neural networks (DNNs) have been increasingly used in safety-critical systems for decision-making (e.g., medical applications, autonomous vehicles, etc.). Early work with adversarial examples shows that they can hinder a DNN’s ability to correctly process inputs. These techniques largely focused on identifying individual causes of failure, emphasizing that minimal perturbations can cause a DNN to fail. Recent research suggests that diverse adversarial examples, those that cause different erroneous model behaviors, can better assess the robustness of DNNs. These techniques, however, use white-box approaches to exhaustively search for diverse DNN misbehaviors. This paper proposes a black-box, model and data-agnostic approach to generate diverse sets of adversarial examples, where each set corresponds to one type of model misbehavior (e.g., misclassification of image to a specific label), thus termed a failure category. Furthermore, each failure category comprises a diverse set of perturbations, all of which produce the same type of model misbehavior. As such, this work provides both breadth and depth-based information for addressing robustness with respect to multiple categories of failures due to adversarial examples. We illustrate our approach by applying it to popular image classification datasets using different image classification models.