基于区块链的工业物联网隐私保护协作学习系统  

Blockchain based Industrial Internet of Things privacyprotection collaborative learning system

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&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：《计算机应用研究》2024年第8期2270-2276,共7页Application Research of Computers

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

摘　　要：为了在保护数据隐私的前提下,充分利用异构的工业物联网节点数据训练高精度模型,提出了一种基于区块链的隐私保护两阶段协作学习系统。首先,使用分组联邦学习框架,根据参与节点的算力将其划分为不同组,每组通过联邦学习训练一个适合其算力的全局模型;其次,引入分割学习,使节点能够与移动边缘计算服务器协作训练更大规模的模型,并采用差分隐私技术进一步保护数据隐私,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,保护模型安全;最后,为了结合多个异构全局模型的优点并进一步提高模型精度,使用每个全局模型的特征提取器从用户数据中提取特征,并将这些特征用作训练集训练更高精度的复杂模型。实验结果表明,该系统在Fashion-MNIST和CIFAR-10数据集上的性能优于传统联邦学习的性能,能够应用于工业物联网场景中以获得高精度模型。To make full use of heterogeneous nodes data from IIoT to train high-accuracy models while protecting data privacy,this paper proposed a privacy-preserving two-stage collaborative learning system based on blockchain.Firstly,it used a grouped federated learning framework to divide participating nodes into different groups based on their computing power.Each group trained a global model suitable for its computing power through federated learning.Secondly,it introduced split learning to enable nodes to collaborate with mobile edge computing servers to train a larger scale model,and used differential privacy technology to further protect data privacy.It stored trained models on the blockchain,and used the consensus algorithm of the blockchain to further prevent attacks from malicious nodes and protect the security of the model.Finally,to combine the advan-tages of multiple heterogeneous global models and further improve model accuracy,it used the feature extractor of each global model to extract features from user data,and used these features as training datasets to train a higher accuracy complex model.Experimental results show that the performance of system on Fashion-MNIST and CIFAR-10 datasets is better than the performance of traditional federated learning.It is suitable for obtaining high-accuracy models in IIoT scenarios.

关 键 词：区块链 工业物联网 隐私保护 协作学习 联邦学习 分割学习 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]