基于区块链的工业物联网隐私保护的异构联邦集成学习系统  

The Ensemble Federated Learning System of Privacy Protection in Industrial Internet of Things Based on Blockchain

作　　者：林峰斌 王灿 吴秋新[1] 李涵[1] 秦宇[2] 龚钢军 LIN Fengbin;WANG Can;WU Qiuxin;LI Han;QIN Yu;GONG Gangjun(School of Applied Science,Beijing Information Science and Technology University,Beijing 100192,China;Institute of Software,Chinese Academy of Sciences,Beijing 100190,China;Beijing Engineering Research Center of Energy Electric Power Information Security,North China Electric Power University,Beijing 102206,China)

机构地区：[1]北京信息科技大学理学院,北京100192 [2]中国科学院软件研究所,北京100190 [3]华北电力大学北京市能源电力信息安全工程技术研究中心,北京102206

出　　处：《江西师范大学学报（自然科学版）》2024年第4期371-382,共12页Journal of Jiangxi Normal University(Natural Science Edition)

基　　金：国家自然科学基金(61604014);国家重点研发计划(2022YFB3105102);未来区块链与隐私计算高精尖中心课题(202203)资助项目.

摘　　要：在工业物联网设备中的数据蕴含着具有重要价值的信息,结合联邦学习技术能够在保护数据隐私的前提下训练模型发掘数据信息.然而,在实际应用中联邦学习仍面临节点异构、数据隐私泄露、服务器单点故障等挑战,为此,该文提出了一种基于区块链的隐私保护异构联邦集成学习系统.首先,采用集成学习对异构联邦学习进行优化,进一步提升了预测的准确率;其次,引入了分割学习使节点能够与移动边缘计算服务器共同完成模型训练,并采用差分隐私技术来进一步保护数据隐私;最后,将训练好的模型存储在区块链上,通过区块链的共识算法进一步防止恶意节点的攻击,保护模型的安全.实验结果表明:使用集成学习优化的方案能够提升异构联邦学习的预测精度,并在节点异构场景下的表现优于传统联邦学习算法.The data in industrial internet of things(IoT)devices contain valuable information.Federated learning technology can be used to train models and discover data information while protecting data privacy.However,practical applications of federated learning still face challenges such as node heterogeneity,data privacy leakage,and single point of failure on the server.Therefore,the privacy-preserving heterogeneous ensemble federated learning system based on blockchain is proposed.Firstly,the system uses ensemble learning to optimize heterogeneous federated learning,further improving the prediction accuracy.Seconelly,it introduces split learning,allowing nodes to jointly complete model training with mobile edge computing servers,and uses differential privacy technology to further protect data privacy.Finally,the trained model is stored on the blockchain,and the consensus algorithm of the blockchain further prevents attacks from malicious nodes,protecting the security of the model.Experiments show that the ensemble learning-optimized scheme can improve the prediction accuracy of heterogeneous federated learning and outperform traditional federated learning algorithms in heterogeneous node scenarios.

关 键 词：区块链 工业物联网 差分隐私 联邦学习 分割学习 集成学习 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]