重庆理工大学学报（自然科学版）2025，Vol.39Issue(3)：1-10,10.DOI:10.3969/j.issn.1674-8425(z).2025.02.001
异质车联网数据的群联邦迁移学习共享方法研究
Research on shared method for swarm federated transfer learning of heterogeneous internet of vehicles
摘要
Abstract
The Internet of Vehicles(IoV)have become an integral part of intelligent transportation systems,offering significant advantages in traffic management,safety,and efficiency.However,the rapid development of IoV also poses big challenges,particularly in terms of data heterogeneity and resource constraints.The diverse types of vehicles,sensors,and driving environments generate highly heterogeneous data,which traditional federated learning methods struggle to handle effectively.These methods often suffer unstable training accuracy and performance degradation due to varied data distributions and limited communication and computational resources available in vehicular networks.To address these challenges,we propose a novel Swarm Federated Transfer Learning(SFTL)method,which optimizes the training process for different data types and scales while enhancing communication efficiency. In our study,three innovations are made. The first in the SFTL method is the consensus device group division mechanism,which is based on the Gaussian Mixture Model(GMM).The GMM is a probabilistic model that assumes all the data points are generated from a mixture of several Gaussian distributions with unknown parameters.By modeling the data distribution,the GMM dynamically identifies potential cluster structures within the data,enabling the segmentation of devices into consensus groups based on their data characteristics.This mechanism effectively manages and analyzes heterogeneous data by grouping devices with similar data distributions together.The GMM-based clustering ensures each group contributes meaningfully to the overall learning process,thereby addressing the inherent heterogeneity of data in the IoV. The second is the Swarm Learning(SL)training mechanism.It leverages blockchain technology to form a decentralized swarm network at the roadside unit level.Each consensus group trains its local model independently within this swarm network.The decentralized nature of the swarm network ensures robustness and security,as it eliminates the single point of failure associated with a central server.The SL mechanism enhances the collaborative learning process among similar device groups by allowing them to share and aggregate their local model updates through a blockchain-based consensus protocol.This not only improves communication efficiency but also addresses issues such as gradient leakage,which is a common problem in traditional federated learning methods.The decentralized aggregation process ensures each device group's contributions are accurately reflected in the global model,achieving more accurate and robust model updates. The third is the inter-group transfer learning mechanism.It uses model pre-training to incrementally migrate information between different consensus device groups.Specifically,it transfers knowledge from data-rich,high-performance groups to those with limited data.This is achieved by pre-training a model on the data-rich group and then fine-tuning it on the data-limited group.The transfer learning approach minimizes model differences and enhances the overall accuracy of federated model aggregation.By leveraging the pre-trained models,the data-limited groups benefit from the knowledge learned by the data-rich groups,thereby improving their local model accuracy.This mechanism not only improves the accuracy of the models but also enhances the generalization capability of the entire system,making it more adaptable to diverse vehicular environments. Experiments conducted on public datasets demonstrate the SFTL method markedly improves training accuracy and communication efficiency compared to baseline methods.Specifically,the model training accuracy is improved by an average of 7％while the communication time is reduced by an average of 10％.Our SFTL method effectively addresses the challenges posed by data heterogeneity in the IoV.It optimizes the communication process by reducing the number of required communication links,making it particularly suitable for large-scale IoV applications where devices are widely distributed and resources are limited.The SFTL method's ability to dynamically segment devices into consensus groups based on their data characteristics ensures each group contributes effectively to the learning process.The decentralized swarm learning mechanism enhances communication efficiency and security,while the inter-group transfer learning mechanism ensures knowledge is effectively shared across different groups.This combination of techniques creates a robust and efficient learning framework well-suited to the complex and dynamic environment of the IoV. Future work will focus on expanding the scale of participating devices and simulating the robustness of the SFTL method under attack scenarios to further validate its practicality and security.We aim to evaluate the method's performance in more realistic and challenging environments,ensuring its applicability in real-world IoV applications.Additionally,the potential for further optimization of the communication and computation processes to enhance the overall efficiency of the SFTL framework will be further explored.关键词
蜂群学习/联邦学习/车联网/高斯混合模型/迁移学习Key words
swarm learning/federated learning/internet of vehicles/Gaussian mixture model/transfer learning分类
信息技术与安全科学引用本文复制引用
康海燕,柯慧敏,邱晓英..异质车联网数据的群联邦迁移学习共享方法研究[J].重庆理工大学学报（自然科学版）,2025,39(3):1-10,10.基金项目
国家社会科学基金项目(21BTQ079) （21BTQ079）
未来区块链与隐私计算高精尖创新中心基金项目(GJJ-23-001) （GJJ-23-001）
北京市教育委员会科技计划项目(KM202011232022) （KM202011232022）
