Federated Learning-Based Interference Modeling for Vehicular Dynamic Spectrum Access

TL;DR

This paper introduces a federated learning-based Radio Environment Map system for vehicular spectrum access, enabling accurate interference modeling and channel prediction in dynamic environments with minimal control data exchange.

Contribution

It proposes a two-layered REM architecture using federated learning to improve interference modeling for vehicular spectrum access, with local and global models merging for enhanced accuracy.

Findings

Effective interference prediction in vehicular environments

Reduced control channel capacity through model parameter exchange

Robustness of local REMs during connectivity disruptions

Abstract

A platoon-based driving is a technology allowing vehicles to follow each other at close distances to, e.g., save fuel. However, it requires reliable wireless communications to adjust their speeds. Recent studies have shown that the frequency band dedicated for vehicle-to-vehicle communications can be too busy for intra-platoon communications. Thus it is reasonable to use additional spectrum resources, of low occupancy, i.e., secondary spectrum channels. The challenge is to model the interference in those channels to enable proper channel selection. In this paper, we propose a two-layered Radio Environment Map (REM) that aims at providing platoons with accurate location-dependent interference models by using the Federated Learning approach. Each platoon is equipped with a Local REM that is updated on the basis of raw interference samples and previous interference model stored in the Global REM. The model in global REM is obtained by merging models reported by platoons. The nodes exchange only parameters of interference models, reducing the required control channel capacity. Moreover, in the proposed architecture platoon can utilize Local REM to predict channel occupancy, even when the connection to the Global REM is temporarily unavailable. The proposed system is validated via computer simulations considering non-trivial interference patterns.