Abstract Adversarial example generation helps to determine the robustness of any given Artificial Neural Network(ANN) dependent system, especially perception-dependent systems due to their large and complex input space, such as those with LiDAR or camera sensors. We develop a new attack based on differentiable rendering to find 3D adversarial inputs that mislead an Artificial Neural Network commonly used in robotics. Our generated inputs are in the form of physically realizable 3D objects. Previous work either generated objects that were not necessarily feasible or relied on hand-crafted rules. We specifically attack a neural network that predicts the quality of robotic parallel-jaw grasps. We optimize the shape of existing 3D objects such that the model makes an incorrect prediction of quality of a particular grasp on that object. Our attack is built on both a novel differentiable implementation of the physics-based Canny-Ferrari grasp quality and a novel self-collision avoidance strategy. With a thorough ablation study over grasps on 79 individual objects, we demonstrate our attack generates the strongest adversarial objects (i.e., makes the model prediction furthest from ground truth).
Mitchell et al. (Thu,) studied this question.