IPRPAS: A Dataset of Physical Adversarial Samples for Assessing Object Detection in Intelligent Vehicles

Physical adversarial samples (PASs) can deceive object detection models (ODMs). Research on the real-world impacts of these samples on ODMs in intelligent vehicles and securing them requires a publicly available benchmark dataset. In this article, we leverage our autonomous driving platform, equipped with a high-resolution video camera and a light detection and ranging, to create a novel and challenging benchmark for securing 2-D and multisensor ODMs. We present images and point clouds of resembling PASs (IPRPASs), a collection of 3409 annotated paired images and point clouds of printed, displayed, or projected resembling PASs (RPASs) in both indoor and outdoor scenarios. IPRPAS includes various common driving objects, such as cars, traffic signs, traffic lights, and persons, under different weather conditions, positions, distances, lighting conditions, and viewing angles. Our experimental evaluation demonstrates that IPRPAS samples deceive state-of-the-art ODMs. Our benchmark is available online at: https://borealisdata.ca/dataset.xhtml?persistentId=doi:10.5683/SP3/O8DLSU. This dataset and the described experiments provide a benchmark for future research on the effectiveness of RPASs and defenses against RPASs in real-world environments.