Excellence in Research: Towards Secure Unmanned Aerial Vehicles-based Systems

该项目旨在增强基于深度学习的无人机（UAV）导航系统对新型攻击向量的弹性，这些攻击向量试图通过对感知数据进行微小更改来误导UAV。 这些系统容易受到不同的攻击，这可能导致无人机的错误引导或黑客攻击，造成重大的安全风险。将考虑不同类型的对抗性攻击。此外，该项目将使用最先进的方法和实验来研究针对这些类型的隐形攻击的防御方法。模型的性能将根据其抵抗攻击的能力来判断，通过模拟和使用无人机试验台的真实世界实验。这项研究的结果将对开发安全可靠的无人机导航系统产生重大影响，并在农作物监测、搜索和救援以及基础设施检查等各个领域具有潜在的应用前景。该项目符合NSF的使命，即促进基础研究，促进国家利益，并通过出版物和开源软件分享其发现，为更广泛的科学界做出贡献。该研究将支持技术开发和北卡罗来纳州AT州立大学代表性不足的研究生和本科生的参与。该项目将研究结合联合收割机两种触发器（即特洛伊木马和对抗性触发器）的攻击场景，以创建一种恶意行为，即在没有特洛伊木马的情况下，模型在对抗性攻击下保持健壮，但在特洛伊木马存在时表现出恶意行为。为了解决这些攻击，将开发一个基于深度学习的检测模型，从对抗性扰动和反事实归因中学习特征，以检测和减轻这些攻击。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This project aims to enhance the resilience of deep learning-based Unmanned Aerial Vehicles (UAV) navigation systems to novel attack vectors that attempt to misguide the UAV by making small changes to the sensed data. Such systems are vulnerable to different attacks, which can result in the misguidance or hacking of the UAV, posing significant safety risks. Different types of adversarial attacks will be considered. Additionally, the project will investigate defense methods against these types of stealthy attacks using state-of-the-art methods as well as experimentations. The performance of the models will be judged according to their ability to resist attacks through simulations and real-world experiments using a UAV testbed. The results of this research will have significant implications for the development of safe and reliable UAV navigation systems, with potential applications in various fields such as crop monitoring, search and rescue, and infrastructure inspection. The project aligns with NSF's mission to promote fundamental research, advance the national interest, and contribute to the broader scientific community by sharing its findings through publications and open-source software. The research will support technical development and engagement of underrepresented graduate and undergraduate students at North Carolina A&T State University. The funding will also enhance development of undergraduate/graduate-level course modules and certificates in autonomy.The project will investigate attack scenarios that combine two triggers, namely trojan and adversarial triggers, to create a malign behavior where the model remains robust in the absence of a trojan under adversarial attacks but behaves maliciously when the trojan is present. To address these attacks, a deep learning-based detection model will be developed that learns features from both adversarial perturbations and counterfactual attributions, to detect and mitigate these attacks. Secure deployment will also be investigated, including remote controller and secure hardware.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.