联邦学习框架下的大数据隐私保护算法研究  

Research on Big Data Privacy Protection Based on Federated Learning Framework

作　　者：颜廷法[1] YAN Tingfa(Basic Teaching Department,Yantai Vocational College,Yantai 264670,China)

机构地区：[1]烟台职业学院基础教学部,山东烟台264670

出　　处：《新乡学院学报》2025年第3期53-57,67,共6页Journal of Xinxiang University

基　　金：山东省教育科学研究项目(24SC126)。

摘　　要：在当前大数据时代,数据隐私保护问题日益凸显,对信息安全构成了严峻挑战。针对大数据环境下的数据隐私泄露问题,提出了一种创新的联邦学习框架下的隐私保护算法。通过Shamir门限秘密共享方案,实现了模型参数的安全共享,同时结合横向与纵向联邦学习策略,构建了一个既灵活又高效的隐私保护模型。在模型训练过程中,采用随机梯度下降算法对本地模型参数进行更新,并通过加密协议对更新的梯度进行加密,以保障数据在传输过程中的安全性。为了进一步提升隐私保护的强度,该算法还集成了差分隐私技术,通过添加适量的随机噪声,有效防止了数据的敏感信息泄露。实验结果表明通过在MedMNIST2数据集上的验证,该算法在降低信息损失率方面表现出色,证明了其在大数据隐私保护方面的实际应用潜力和效果。In the current era of big data,issues surrounding data privacy protection are becoming increasingly prominent,posing significant challenges to information security.This paper addresses the problem of data privacy leakage in the big data environment by proposing an innovative privacy protection algorithm based on federated learning framework.Through the Shamir’s secret sharing scheme,the algorithm achieves secure sharing of model parameters and constructs a flexible and efficient privacy protection model by combining horizontal and vertical federated learning strategies.During the model training process,local model parameters are updated using the stochastic gradient descent algorithm,and the updated gradients are encrypted using encryption protocols to ensure data security during transmission.To further enhance privacy protection,the algorithm integrates differential privacy techniques,effectively preventing sensitive information leakage by adding appropriate amounts of random noise.Experimental results on the MedMNIST2 dataset demonstrate outstanding performance in reducing information loss rates,confirming its practical potential and effectiveness in big data privacy protection.

关 键 词：大数据隐私 联邦学习 加密协议 差分隐私 模型参数更新 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]