Adversarial Neon Beam: A light-based physical attack to DNNs

Abstract

In the physical world, the interplay of light and shadow can significantly impact the performance of deep neural networks (DNNs), leading to substantial consequences, as exemplified by incidents such as the Tesla self-driving car collision caused by an unexpected light flash. Traditional methods involving stickers for physical attacks have inherent limitations, particularly in terms of stealthiness. In response, researchers have delved into light-based perturbations, including lasers and projectors, with the objective of achieving stealthy attacks. However, these efforts have often fallen short in terms of achieving robustness.In our study, we introduce a pioneering black-box light-based physical attack known as Adversarial Neon Beam (AdvNB). Our method stands out for its excellence in attack modeling, efficient attack simulation, and robust optimization, striking a harmonious balance between robustness and efficiency. We employ effectiveness, stealthiness, and robustness as the key metrics to evaluate the proposed AdvNB. Through rigorous evaluation, we attain an impressive 84.40% attack success rate in digital attacks, requiring an average of 189.70 queries. In real-world scenarios, our method excels with a 100% attack success rate indoors and a commendable 81.82% success rate outdoors. AdvNB demonstrates its stealthiness through comparisons with baseline samples, and it further underscores its robustness by consistently achieving a success rate exceeding 80% when targeting advanced DNN models.We carry out a comprehensive analysis of the proposed attack and note that the generated perturbations share similarities with objects present in the dataset or real-world settings. Additionally, we implement adversarial defense mechanisms against AdvNB. Given its superior performance compared to baseline methods as a light-based attack, we advocate for its broader acknowledgment and recommend its adoption as a reference point for future research and practical applications. Our code and data can be accessed from the following link:https://github.com/ChengYinHu/AdvNB.