Adversarial Attacks on Deep Learning Models of Computer Vision: A Survey

Deep learning plays a significant role in academic and commercial fields. However, deep neural networks are vulnerable to adversarial attacks, which limits its applications in safety-critical areas, such as autonomous driving, surveillance, and drones and robotics. Due to the rapid development of adversarial examples in computer vision, many novel and interesting adversarial attacks are not covered by existing surveys and could not be categorized according to existing taxonomies. In this paper, we present an improved taxonomy for adversarial attacks, which subsumes existing taxonomies, and investigate and summarize the latest attacks in computer vision comprehensively with respect to the improved taxonomy. Finally, We also discuss some potential research directions.