面向分类任务的隐私保护协作学习技术  

Privacy-preserving collaborative learning technology for classification

作　　者：黎兰兰 张信明[1] Li Lanlan;Zhang Xinming(School of Computer Science and Technology,University of Science and Technology of China,Hefei 230026,China)

机构地区：[1]中国科学技术大学计算机学院,安徽合肥230026

出　　处：《网络安全与数据治理》2023年第5期36-43,共8页CYBER SECURITY AND DATA GOVERNANCE

基　　金：国家重点研发计划(2020YFB2103803)。

摘　　要：随着相关法律法规的发布和人们隐私意识的觉醒,隐私保护要求不断提高。针对分类任务场景,提出了一种隐私性与效用性并重的面向分类任务的隐私保护协作技术(PCTC)。在隐私安全方面,将同态加密和差分隐私相结合,并设计了一种安全聚合机制,实现更加健壮的隐私保护;在效用性方面,引入信息熵的计算因子对标签可信度进行度量,降低标注噪声对模型性能的影响。最后对所提出的方案进行了安全性分析,并在公开数据集上进行了实验验证。结果表明PCTC在保证数据隐私安全的同时大幅度提升了模型性能,具有较为广阔的应用前景。With the release of relevant laws and regulations and the awakening of people’s privacy awareness,the requirements for privacy protection are constantly increasing.Aiming at the scenario of classification,this paper proposes a Privacy-preserving Collaborative Learning Technology for Classification(PCTC)that emphasizes both privacy and utility.In terms of privacy,homomorphic encryption and differential privacy are combined and a secure aggregation mechanism is designed to achieve more robust privacy protection.In terms of utility,the calculation factor of information entropy is introduced to measure the credibility of labels,which can reduce the impact of labeling noise on model performance.Finally,the security analysis of the proposed scheme is carried out,and the experiments are implemented on public datasets.The results show that PCTC significantly improves model performance while ensuring privacy and security of the data,and has broad application prospects.

关 键 词：隐私保护 数据标注 分类任务 同态加密 差分隐私 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]