去中心跨孤岛的混合联邦学习通信算法研究

Hierarchical Communication in Decentralized and Cross-Silo Federated Learning

为解决异构数据孤岛之间难以开展安全机器学习的问题,提出了一种异构数据孤岛之间的联邦学习通信方式,实现了横向和纵向混合的联邦学习通信,突破了传统联邦学习横向和纵向参与方之间模型结构不统一的通信壁垒。基于政府、银行等机构的特殊性隐私需求,在混合联邦学习模型的基础上进一步去除了第三方聚合器,计算只在参与方之间进行,大大提高了本地数据的隐私安全性。同时针对上述模型中纵向同态加密为通信过程带来的计算速度瓶颈问题,通过增加本地迭代轮次q将纵向联邦学习的加密时间缩短了10倍以上,降低了横向参与方与纵向参与方间的计算瓶颈,并且精度损失不超过5%。

Federated learning has become increasingly important for modern machine learning,especially for data privacy sensitive scenarios.It is difficult to carry out secure machine learning between heterogeneous data islands.A federated learning communication mode between heterogeneous data islands is proposed,which realizes the hybrid federated learning communication between horizontal and vertical,and breaks the communication barrier of the disunity of model structure between horizontal and vertical participants in traditional federated learning.Based on the special privacy requirements of the government,banks and other institutions,the third party aggregator is further removed on the basis of the hybrid federated learning model,and the calculation is carried out only among the participants,which greatly improves the privacy security of local data.In view of the computational speed bottleneck caused by vertical homomorphic encryption in the communication process in the above model,by increasing the local iteration round q,the encryption time of vertical federation learning is shortened by more than 10 times,and the computational bottleneck between horizontal and vertical participants is reduced,and the accuracy loss is less than 5%.