A Compressed Decentralized Federated Learning Framework for Enhanced Environmental Awareness in V2V Networks

Decentralized Federated Learning (FL) policies have been recently studied in vehicular networks for learning driving-related information from distributed datasets without sharing privacy-sensitive data. Despite their outstanding learning performances, they generally suffer from significant communication overhead, making them unsuitable for latency-sensitive vehicular applications. In this paper, we propose a general framework of communication-efficient accelerated FL policies tailored for decentralized vehicular setups. The proposed framework makes use of adaptive gradient methods jointly with compressed updates to reduce communication overhead and speed up the learning process. A challenging automotive use case is used to evaluate the developed framework, where vehicles aim at obtaining an improved environmental awareness via semantic segmentation. Numerical results show that the proposed FL framework improves over state-of-the-art baselines both in terms of communication efficiency and learning performances.