基于BCP的联合委托学习模型及协议  

BCP-based joint delegation learning model and protocol

作　　者：高胜 向康[1,3] 田有亮 谭伟杰[1] 冯涛 吴晓雪 GAO Sheng;XIANG Kang;TIAN Youliang;TAN Weijie;FENG Tao;WU Xiaoxue(State Key Laboratory of Public Big Data,College of Computer Science and Technology,Guizhou University,Guiyang 550025,China;School of Information,Central University of Finance and Economics,Beijing 100081,China;Institute of Cryptography and Data Security,Guizhou University,Guiyang 550025,China;School of Computer and Communication,Lanzhou University of Technology,Lanzhou 730050,China;Institute for Metrology and Calibration of Guizhou,Guiyang 550000,China)

机构地区：[1]贵州大学计算机科学与技术学院公共大数据国家重点实验室,贵州贵阳550025 [2]中央财经大学信息学院,北京100081 [3]贵州大学密码学与数据安全研究所,贵州贵阳550025 [4]兰州理工大学计算机与通信学院,甘肃兰州730050 [5]贵州省计量测试院,贵州贵阳550000

出　　处：《通信学报》2021年第5期137-148,共12页Journal on Communications

基　　金：国家自然科学基金资助项目(No.61662009,No.61772008,No.U1836205,No.62072487);贵州省科技重大专项计划基金资助项目(No.20183001);贵州省科技计划基金资助项目(黔科合基础[2019]1098,黔科合平台人才[2019]5703);贵州省高层次创新型人才基金资助项目(黔科合平台人才[2020]6008);贵州大学引进人才科研基金资助项目(贵大人基合字[2020]61);贵州大学培育基金资助项目(贵大培育[2019]56)。

摘　　要：为了实现数据安全共享的同时减少客户端在数据挖掘过程中的计算成本,基于BCP同态加密算法提出了联合委托学习模型及协议。首先,针对决策树模型的安全构造提出了基于虚假记录的隐私保护方法。其次,根据数据垂直分布与水平分布的情况,基于隐私保护委托点积算法和隐私保护委托求熵算法提出了相应的委托学习协议。最后,给出了委托学习协议及决策树模型结构的安全性证明和性能分析。结果表明,基于虚假记录的隐私保护方法不会影响最终模型的构建,并且各客户端最终获得的模型与真实数据构建的模型具有一致性。In order to realize data security sharing and reduce the computing costs of clients in data mining process,a joint delegation learning model and protocol based on BCP homomorphic encryption algorithm was proposed.Firstly,a privacy preserving method based on false records was proposed for the security of decision tree model.Secondly,in view of the vertical and horizontal distribution of data,the corresponding delegation learning protocols based on privacy preserving delegation dot product algorithm and privacy preserving delegation entropy algorithm was proposed.Finally,the security proof and the performance analysis of delegation learning protocols and decision tree model structure were given.The results show that the privacy protection method based on false records does not affect the final model construction,and the final model obtained by each client is the same as that constructed by real data.

关 键 词：委托学习 决策树 BCP同态加密 数据挖掘 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]