Adaptively Compressed Swarm Learning for Distributed Traffic Prediction over IoV-Web3.0

Xin Song, Lixing Chen, Junhua Tang, Jianhua Li, Yang Bai, Wu Yang

Abstract: Web3.0 presents a promising paradigm to facilitate secured collaborations and decentralized interoperability in the Internet of Vehicles (IoV). This paper proposes a novel decentralized machine learning approach, namely Adaptively Compressed Swarm Learning (ACSL), over IoV-Web3.0, with a primary objective of traffic prediction using Roadside units (RSUs). ACSL systematically characterizes the model exchange process of Swarm Learning over the stamp system of the Ethereum-based swarm network, and formulates a practical stamp-constrained learning environment for IoV operators over web3.0. ACSL works under time-varying stamp budgets and judiciously determines the compression rate of model parameters to strike a balance between stamp cost (i.e., blockchain cost) and the accuracy of traffic prediction. ACSL is realized via the adaptive Top-K mechanism that only exchanges the K largest model parameters. Particularly, it designs a dynamic selection strategy for the value of K that maximizes the learning performance under hard stamp budgets. We evaluate the proposed method on a real-world PeMS dataset, the experimental results demonstrate that ACSL reduces the stamp cost by approximately 10x and the training latency by 5.3x compared to vanilla SL, and provides an improvement of 6.96% in traffic prediction accuracy compared to classical Top-K algorithm.