Federated Machine Learning in Network Environments for Connected and Automated Mobility Applications

Abstract

5G networks and beyond (B5G) are expected to transform mobile communications, enabling the seamless integration of people, devices, and sensors (Internet of Things - IoT) within cyber-physical environments, ultimately realizing the concept of the "Internet-of-Everything" (IoE). A key enabler of this transformation is the convergence of Cloud Computing and the IoT domain, referred to as the 5G network and compute continuum. This continuum facilitates data exchange, processing, and decision-making across all 5G domains, from the cloud and network edge to the IoT domain, known as the Extreme-Edge.

Artificial Intelligence and Machine Learning (AI/ML) play a central role in this evolution. From Industry 5.0 and automotive sectors to infotainment, education, and e-health, AI/ML technologies have introduced innovative solutions that allow computer systems to learn from environmental data, enabling fully autonomous systems capable of decision-making in a human-like manner.

Traditionally, AI/ML in network environments has been implemented in a centralized manner, with data collection and processing occurring in central clouds. However, recent research has shifted focus toward distributed solutions to leverage the data generated by mobile client devices. Unlike Centralized Learning (CL), Distributed Learning (DML) methods, such as Federated Learning (FL), offload computation to client devices, offering benefits such as scalability, cost-efficiency, and privacy preservation for user data.

Existing research on DML primarily focuses on the performance (accuracy) of trained models, often overlooking the practical aspects, such as the impact on underlying network resource consumption. This dissertation seeks to address these gaps by investigating the implementation of DML schemes from a systems perspective. Specifically, it examines the Cooperative, Connected, and Automated Mobility (CCAM) applications in the automotive domain, which have stringent requirements for both network performance (e.g., latency) and application performance (e.g., safety).

To begin, we conduct an end-to-end performance comparison between CL and FL, analyzing training efficiency and resource consumption across all network stakeholders: clients, the network, and cloud/edge infrastructure. We explore the complex issue of ML scheme selection, considering various system parameters and constraints, including network and mobility conditions as well as AI/ML metrics (e.g., convergence). This analysis identifies the trade-offs between critical parameters when choosing between CL and FL.

Next, inspired by ML operations (MLOps) and continuous learning, we examine how concept drift—changes in data distributions over time—affects the performance of distributed ML models, particularly in mobile and vehicular networks like those used in CCAM applications. These networks are highly dynamic and prone to drift. After understanding the impact of concept drift, we propose novel techniques to manage it in a resource-efficient manner.

Finally, we validate our simulation-based findings through real-world testing. We first conduct a large-scale measurement campaign to collect network Quality-of-Service (QoS) and mobility data, which is then used to demonstrate a practical FL application: distributed QoS prediction. Our framework is extended to manage and orchestrate multiple FL services, introducing an orchestrator for Extreme-Edge and IoT devices performing AI/ML tasks. This orchestrator is deployed on a commercial-grade 5G testbed and evaluated using both mobile (in-vehicle) and static (lab-based) devices. The results demonstrate the feasibility of lifecycle management for multiple services, particularly in the automotive sector, showcasing the potential of FL in large-scale environments.