LAUSR

Unmanned aerial vehicles (UAVs) have been widely adopted to assist infrastructure inspection tasks, since they have shown their potential to benefit the inspection in terms of efficiency, cost, and safety. To improve the effectiveness of the inspection, there has been a growing focus of recent research in integrating AI techniques into UAV infrastructure inspection and has achieved promising results. However, no prior work has studied whether such integration will also introduce new security concerns, especially considering the existence of potential vulnerabilities in underlying AI models toward adversarial inputs. In this article, we perform the first study to fill this critical gap by identifying and validating the security vulnerabilities of AI models in the context of UAV infrastructure inspection with a focus on bridge infrastructure. To understand the security property of AI-assisted UAV bridge inspection, we design a two-stage approach for the construction of effective adversarial inputs, with which we successfully validate the existence of security vulnerability using dynamic analysis. Spatial constraints, physical limits, and dynamic environmental changes are taken into consideration in our analysis to make it practical in the physical world. Our evaluation results on a real-world data set show that our constructed adversarial inputs can mislead the UAV to miss detecting a significant amount of risk-prone regions by exploiting the identified vulnerability. Based on such an observation, this article also discusses the defenses based on adversarial training to improve the robustness of AI-assisted UAV bridge inspection, which has been demonstrated to be effective according to our experimental evaluation results.