Vehicular Intelligence at the Edge: A Decentralized Federated Learning Approach for Technology Recognition

Abstract

In the evolving landscape of vehicular networks, the need for robust, scalable, and decentralized learning mechanisms is paramount. This paper introduces a novel Decentralized Federated Learning (DFL) framework for wireless technology recognition in vehicular networks, essential for intelligently allocating spectrum resources in multi-Radio Access Technology (multi-RAT) scenarios. In contrast with centralized learning at the base station level, our approach leverages Roadside Units (RSUs) for model training and aggregation, eliminating central server dependency and enhancing resilience to single points of failure. Each vehicle trains a Convolutional Neural Network (CNN) for wireless technology recognition using the Fourier transform of In-phase and Quadrature (IQ) samples collected from a specific combination of technologies. The proposed framework is comprised of two steps. First, Centralized Federated Learning (CFL) is employed at the RSU level to create an aggregated model, considering the users' connectivity status. Second, DFL is utilized to establish a global model at each RSU by sharing models with neighboring RSUs. This approach not only preserves data privacy and security but also optimizes learning by leveraging local computations and minimizing the need for extensive data transmission. Our experimental analysis validates the viability of this approach in providing a scalable and resilient solution for technology recognition in vehicular networks. Our results indicate that DFL surpasses its centralized counterpart by 30% in sparse deployments with low connectivity rates.