A Little Fog for a Large Turn

A Little Fog for a Large T urn Harshitha Machiraju, Vineeth N Balasubramanian Indian Institute of Technology, Hyderabad, India {ee14btech11011, vineethnb }@iith.ac.in Abstract Small, carefully crafted perturbations called adversarial perturbations can easily fool neural networks. However, these perturbations are largely additive and not naturally found. W e turn our attention to the field of Autonomous navigation wherein adverse weather conditions such as fog have a drastic effect on the predictions of these systems. These weather conditions are capable of acting like natural adversaries that can help in testing models. T o this end, we introduce a general notion of adversarial perturbations, which can be created using generative models and provide a methodology inspired by Cycle-Consistent Generative Adversarial Networks to generate adversarial weather conditions for a given image. Our formulation and results show that these images provide a suitable testbed for steering models used in Autonomous navigation models. Our work also presents a more natural and general definition of Adversarial perturbations based on Perceptual Similarity. 1 1. Introduction Autonomous navigation has occupied a central position in the efforts of computer vision researchers in recent years. Autonomous vehicles can not only aid navigation in urban areas but also provide critical support in disaster-affected areas, places with unknown topography (such as Mars), and many more. The vast potential of the applications thereof and the feasibility of the solutions in contemporary times has led to the growth of several organizations across industry, academia, and government institutions that are investing significant efforts on self-driving vehicles.