摘要
纵向联邦学习作为实现“数据可用不可见”的重要技术范式,其核心的学习过程是基于安全求交的样本对齐.已有的安全求交虽然保护了非交集信息的隐私不被泄露,但无法满足交集部分用户ID的隐私保护需求.抽象出一种基于匿踪对齐的全匿踪纵向联邦学习框架,确保联邦学习全链路都不会泄露各持有方集合的隐私信息;提出一种基于多方安全计算的框架实现方法,在保持全匿踪的条件下进行联合建模,迭代训练直到模型收敛;通过实验验证了该框架的高性能与低误差特性,能够较好地应用于实践.
As a key technical paradigm to achieve“data availability and invisibility”,the core process of vertical federated learning is sample alignment based on private set intersection.Although the private set intersection protects the privacy of non-intersected information,it can’t meet the privacy protection requirements of user IDs in the intersected set.This paper proposes a fully anonymous vertical federated learning framework based on anonymous alignment to ensure that no private information of each holder set will be disclosed during the whole process.An implementation framework based on secure multi-party computation is proposed for fully anonymous joint modeling.The high performance and low error characteristics of the framework are verified through experiments,indicating it can be better applied in practice.
作者
尤志强
李月
姜玮
方竞
陈立峰
卞阳
You Zhiqiang;Li Yue;Jiang Wei;Fang Jing;Chen Lifeng;Bian Yang(Shanghai Fudata Technology Co.,Ltd.,Shanghai 200120;China EPort Data Center,Shanghai Branch,Shanghai 200120;Science&Technology Division of Shanghai Customs,Shanghai 200135)
出处
《信息安全研究》
CSCD
北大核心
2024年第6期506-512,共7页
Journal of Information Security Research
关键词
纵向联邦学习
安全求交
匿踪求交
多方安全计算
匿踪学习
vertical federated learning
private set intersection
anonymous alignment
secure multi-party computation
anonymous federated learning
