基于自动化特征组合的隐私保护风险识别机制  

Privacy protection risk identification mechanism based on automated feature combination

作　　者：蔡民超 姚宏伟 王旸[1] 秦湛 陈少梦 任奎[1] CAI Minchao;YAO Hongwei;WANG Yang;QIN Zhan;CHEN Shaomeng;REN Kui(School of Cyber Science and Technology,Zhejiang University,Hangzhou 310007,China;Hangzhou Kuaidi Technology Co.,Ltd.,Hangzhou 310000,China)

机构地区：[1]浙江大学网络空间安全学院,浙江杭州310007 [2]杭州快迪科技有限公司,浙江杭州310000

出　　处：《通信学报》2024年第11期1-14,共14页Journal on Communications

基　　金：国家重点研发计划基金资助项目(No.2021YFB3100300);国家自然科学基金资助项目(No.U20A20178,No.62072395,No.62206207)。

摘　　要：异常行为识别(AD)算法在实际应用中,通常会面临特征组合优化困难、分类器准确率难提高、模型应用效率低等技术挑战。用户所产生的多维数据具有丰富的空间结构信息,围绕这些多维数据的特点,在通过同态加密的隐私保护方式进行数据脱敏的基础上,针对特征组合优化困难的技术挑战,提出并实现了首个基于特征分箱的自动化特征组合优化模型算法,该算法在特征组合优化方面提升了99.93%的计算效率。基于自动化特征组合优化模型筛选出的重要特征所组合的规则仍存在分类器准确率难提高的技术挑战,故将自动化筛选出的重要特征融入识别模型中,设计并实现了首个规则和算法的交叉应用模型,并将该方式应用到基于用户多维信息的异常行为识别中,在识别先享不付类异常用户的具体场景中实现资金挽损效率提升27.78%。In practice,the anomaly detection(AD)algorithm usually faced technical challenges such as difficulty in opti‐mizing feature combinations,difficulty in improving classifier accuracy,and low model application efficiency.The multi‐dimensional data generated by users was with rich spatial structure information,revolved around the characteristics of the multidimensional data.Building upon the privacy protection method using homomorphic encryption,the technical challenge of optimizing feature combinations was addressed.The first automated feature combination optimization model algorithm based on feature binning was proposed and implemented.This algorithm enhanced computational effi‐ciency in feature combination optimization by 99.93%.The rules combined by the important features selected by the au‐tomatic feature combination optimization model still faced the technical challenge of difficulty in improving the classi‐fier accuracy.Therefore,the important features selected automatically were integrated into the recognition model,the first cross-application model of rules and algorithms was designed and implemented.This approach was applied to anomaly detection based on multi-dimensional user data,resulting in a 27.78%increase in funds saved in the specific scenario of identifying abnormal users who enjoy first but do not pay.

关 键 词：异常行为识别 多维信息 风险识别 可信度模型 同态加密 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]