面向智能交通的安全高效个性化联邦学习框架

Secure and efficient personalized federated learning framework for intelligent transportation

下载PDF

导出

摘要区块链与联邦学习的融合为智能交通系统提供了一种去中心化、隐私保护的数据协同方案。然而终端设备类型多样、数据分布呈非独立同分布(Non-IID),且传统单链结构难以满足高并发训练的吞吐与延迟需求。为此,提出一种安全高效的个性化联邦学习框架(SCPFL),利用区块链分片按节点地理区域划分网络,实现分片内并行训练与聚合以提升可扩展性和吞吐率;同时,引入基于余弦相似度的个性化方法,在支持个性化模型生成的同时促进跨分片知识共享;并设计筛选与激励双重机制,以抵御恶意节点并提升设备参与度。实验表明,SCPFL在CIFAR-10、MNIST和交通标志数据集GTSRB上的模型性能均优于基线方法,其中相较主流个性化方法(如Ditto)最高可提升18.3%的精度,收敛速度最快提高50%;在系统效率上,相较传统基于主子链的联邦学习架构,系统吞吐量提高约53%,同时处理时间、CPU和内存消耗分别降低33.3%、17.7%和19.7%。 Blockchain-federated learning(FL)integration enables decentralized and privacy-preserving collaboration in intelligent transportation systems,but device heterogeneity,non-IID data,and single-chain bottlenecks limit scalability and efficiency.This paper proposed a secure and cost-efficient personalized federated learning(SCPFL)framework.By leveraging blockchain sharding,SCPFL supported parallel training and aggregation across regions,while a cosine-similarity-based perso-nalization method balanced individual adaptation and cross-shard knowledge sharing.A dual mechanism of filtering and incentives further enhanced robustness and participation.Experiments on CIFAR-10,MNIST,and German traffic sign recognition benchmark(GTSRB)datasets show that SCPFL improves accuracy by up to 18.3%and accelerates convergence by 50%compared with the state-of-the-art personalized baselines(e.g.,Ditto).In terms of system efficiency,SCPFL achieves 53%higher throughput and reduces processing time,CPU,and memory usage by 33.3%,17.7%,and 19.7%over traditional main-subchain FL architectures.

作者王逸飞胡颍蔡婷陈炜吴雨芯李晓丽叶志伟 Wang Yifei;Hu Ying;Cai Ting;Chen Wei;Wu Yuxin;Li Xiaoli;Ye Zhiwei(School of Computer Science,Hubei University of Technology,Wuhan 430068,China;Hubei Provincial Key Laboratory of Green Intelligent Computing Power Network,Hubei University of Technology,Wuhan 430068,China;School of Computer Science,Hubei University of Arts and Science,Xiangyang Hubei 441053,China)

机构地区湖北工业大学计算机学院湖北工业大学绿色智能算力网络湖北省重点实验室湖北文理学院计算机学院

出处《计算机应用研究》北大核心 2026年第3期712-719,共8页 Application Research of Computers

基金国家自然科学基金资助项目(62302154,U23A20318,62306108,62376089) 湖北省自然科学基金资助项目(2024AFB882) 湖工大博士科研启动基金资助项目(XJ2022006701) 智能感知系统与安全教育部重点实验室开放基金资助项目(KLISSS202404) 湖北省教育厅科研重点项目(D20242602) 湖北省高等学校优秀中青年科技创新团队计划资助项目(T2023007)。

关键词个性化联邦学习区块链分片安全激励机制智能交通系统 personalized federated learning sharding blockchain security incentive mechanism intelligent transportation system(ITS)

分类号 TP393 [自动化与计算机技术—计算机应用技术]