Consensus-Based Decentralized Federated Learning for Model Training Services in IoV

With the rapid development of the Internet of Vehicles (IoV), massive amounts of distributed data are continuously generated, raising critical challenges in ensuring service reliability and security. Federated Learning (FL) has emerged as a promising approach to address these challenges by enabling privacy-preserving and scalable model training services. However, traditional FL frameworks, which rely on a central server for model aggregation, suffer from single-point failures and limited scalability. Furthermore, in IoV environments, the vast amounts of sensor-collected data often exhibit redundancy, making the existing aggregation strategy inapplicable, leading to inefficient model updates and ultimately degrading performance. To address these challenges, we propose C-DFL (Consensus-based Decentralized Federated Learning), a novel decentralized data processing framework specifically designed to enhance model training services in intelligent vehicle environments. The core innovation of C-DFL lies in its ability to transform data features into sketches, which are then distributed among vehicles to calculate non-redundant data. Considering the quantity of nonredundant data, we design a new federated learning aggregation function. We evaluate the performance of C-DFL through comprehensive experiments conducted on an NS3-based simulation platform using two real-world datasets. The results demonstrate that C-DFL consistently outperforms the compared Decentralized Federated Learning (DFL) methods in terms of accuracy and convergence rate.