Federated Object Detection in Vehicular Networks: Optimizing Learning and Communication Performance

NAISS 2024/5-153


NAISS Medium Compute

Principal Investigator:

José Mairton Barros da Silva Júnior


Uppsala universitet

Start Date:


End Date:


Primary Classification:

10202: Information Systems (Social aspects to be 50804)

Secondary Classification:

20204: Telecommunications

Tertiary Classification:

20203: Communication Systems



The fifth generation (5G) and the emerging sixth generation (6G) of cellular networks aim to support vehicular communications, including communication between vehicles, pedestrians, road infrastructure. Vehicular communications often face difficult wireless propagation conditions due to fast-varying wireless communication channels, and have to support low latency, high reliability, and high data rates with the vehicles forming dynamic topologies. With the help of cellular networks, current vehicular applications will use machine learning services to realize the vision of assisted and self-driving systems. The sensors at the vehicles generate large quantities of private-sensitive data, and the vehicles may benefit on the data exchange between the network and themselves. Centralized machine learning uses such a large amount of data to make inference and real-time decision for assisted-driver systems, such as vehicular 2D and 3D object detection. Since the data is distributed among vehicles, distributed learning approaches are recommended. Accordingly, federated learning (FL) is a promising emerging distributed paradigm that pushes learning and computation into the devices to improve communication efficiency and privacy. FL trains a global objective function (learning task) using a central server and a subset of all nodes, which use their own local data to solve a local objective function, through an iterative process between the devices and the central server. Unlike other centralized and distributed machine learning techniques, federated learning is designed for scenarios in which the devices have to send a limited number of messages (communication efficiency), the devices do not share the raw data (privacy), and are heterogeneous in the data and in the computation capabilities (statistical and systems heterogeneity). 5G and 6G support communication demands for assisted and self-driving, while machine learning proposes FL for distributed scenarios. It is only by integrating wireless communications and distributed machine learning domains that the vision of assisted and self-driving will become real. This new ecosystem requires the FL algorithms to address critical constraints from vehicular communications, while resource allocation algorithms adapt wireless communication resources to the FL tasks. Our objective is to explore distributed applications for assisted driving systems that can benefit from vehicular communications. We aim to assess the performance of FL algorithms from various perspectives, including communication efficiency, privacy, and suitability for wireless communication, particularly for critical learning tasks in vehicular applications, such as object detection. The outcomes of our study will contribute to a better understanding of the advantages and challenges associated with employing FL over vehicular networks. The research is being conducted by a team of researchers from my group at Uppsala University, comprising three master's students, one doctoral student, and one postdoctoral researcher. Each member of the team is working on specific aspects of the project: • Master students: o Communication-Efficient Federated Object Detection for Vehicular Applications; o Vertical Federated Object Detection for Vehicular Networks; o Federated Object Detection over Vehicular Communications. • Doctoral student: o Distributed Machine Learning for Next Generation Wireless Communications. • Postdoctoral researcher: o Security and Privacy in Distributed Machine Learning over Wireless Networks.