A Compositional Simulation Framework for Testing Adversarial Robustness of Deep Neural Networks

Deep neural networks (DNN) have reached impressive performance in computer vision, making them a natural choice for object detection problems in automated driving. However, DNNs used for object detection are known to be highly vulnerable to adversarial attacks. Even small changes to the input such as adding a customized noise pattern that remains invisible to the human eye, can stimulate silent prediction errors. In this study, we present a compositional simulation framework for testing the adversarial robustness of DNNs used in object detection. We demonstrate our framework with a comprehensive case study of a speed-sign detector model with two different adversarial attacks.