Collaborative Research: OAC CORE: Federated-Learning-Driven Traffic Event Management for Intelligent Transportation Systems

智能交通系统（ITS）利用智能交通监控和机器学习（ML）技术来优化交通管理和保证驾驶安全。目前，集中式机器学习是ITS的主流学习方法，其中分布式边缘设备（例如智能交通摄像头和行车记录仪）之间的大量交通视频数据被传输到中央服务器以训练机器学习模型，导致效率和隐私问题令人望而却步。联邦学习（FL）是一种很有前途的范例，它利用分布式设备的计算能力，在保持数据本地和安全的同时，支持在大规模数据上对共享ML模型进行协作训练。不幸的是，现有的FL包不能在资源有限的设备上完全支持FL，而这些设备在道路基础设施边缘设备中占主导地位。该项目旨在建立一个边缘友好的网络基础设施，使FL能够以高效、安全和保护隐私的方式部署到ITS应用程序中。该研究将在自然驾驶研究和交通冲突预测等诸多交通应用领域带来革命性的进步。拟议的网络基础设施将用于现实世界的交通管理，以增强交通机构的态势感知和决策能力。拟议的软件工具将是开源的，以增强广泛的ITS社区的研究基础设施。教育活动包括课程开发、学生指导和对K-12学生的拓展。该项目将从效率、安全性和隐私的关键角度建立关于FL的新的理论和实践结果，这是在ITS中大量资源有限的道路基础设施边缘设备上广泛采用和部署所必需的三个属性。具体而言，(1)本项目将系统地研究智能交通系统中不同类型的效率问题（如昂贵的计算成本、高通信消耗和低设备利用率）与FL之间的相互作用。(2)本项目将为ITS中数据和模型的恶意攻击提供经验和认证防御的理论和实践安全工具。(3)本项目将研究交通视频数据中FL与隐私之间的关系，提出新的基于理论的设计和FL架构，如隐私保护数据和模型共享。(4)本项目将使用NVIDIA Jetson Nano器件开发一个真实的分布式测试平台，对上述方法进行测试。这些小型设备可以部署在FL的接线盒和车辆中，以服务于ITS应用。该项目由高级网络基础设施办公室（OAC）核心研究计划和促进竞争研究的既定计划（EPSCoR）共同资助。该奖项反映了美国国家科学基金会的法定使命，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。

Intelligent transportation systems (ITS) utilize smart traffic surveillance and machine learning (ML) technologies to optimize traffic management and guarantee driving safety. Currently, centralized ML is the mainstream learning method in the ITS, where vast amounts of traffic video data among distributed edge devices (e.g., smart traffic cameras and dashcams) are transmitted to a central server to train an ML model, resulting in prohibitive efficiency and privacy concerns. Federated learning (FL) is a promising paradigm that leverages the computing power of distributed devices to enable collaborative training of shared ML models over large-scale data while keeping the data local and safe. Unfortunately, existing FL packages fail to fully support the FL on resource-limited devices, which dominate the road infrastructure edge devices. This project aims to build an edge-friendly cyberinfrastructure that allows FL to be deployed for ITS applications in an efficient, secure, and privacy-preserving manner. The proposed research will bring transformative advances in many transportation applications, such as naturalistic driving study and traffic conflict prediction. The proposed cyberinfrastructure will be deployed for real-world traffic management to enhance transportation agencies’ situational awareness and decision-making capabilities. The proposed software tools will be open source to enhance the research infrastructure for the broad ITS communities. Educational activities include curriculum development, student mentoring, and outreach to K-12 students.The project will establish new theoretical and practical results about the FL from the critical perspectives of efficiency, security, and privacy — three properties necessary for broad adoption and deployment on the massive resource-limited road infrastructure edge devices in the ITS. Specifically, (1) this project will systematically investigate the interplay between the FL and distinct types of efficiency issues in the ITS, such as expensive computation cost, high communication consumption, and low device utilization. (2) This project will provide theoretical and practical security tools for both empirical and certified defenses against malicious attacks on data and models in the ITS. (3) This project will investigate the relationship between FL and privacy in the traffic video data by proposing new theoretically grounded designs and FL architectures, such as privacy-preserving data and model sharing. (4) This project will develop a real distributed testbed with NVIDIA Jetson Nano devices to test the above-proposed methods. These small devices can be deployed in junction boxes and vehicles for FL to serve ITS applications.This project is jointly funded by the Office of Advanced Cyberinfrastructure (OAC) Core Research program and the Established Program to Stimulate Competitive Research (EPSCoR).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.