基于相关性的Swarm联邦降维方法

Swarm Federated Dimensionality Reduction Method Based on Correlation

联邦学习(Federated learning,FL)在解决人工智能(Artificial intelligence,AI)面临的隐私泄露和数据孤岛问题方面具有显著优势.针对联邦学习的已有研究未考虑联邦数据之间的关联性和高维性问题,提出一种基于联邦数据相关性的去中心化联邦降维方法.该方法基于Swarm学习(Swarm learning,SL)思想,通过分离耦合特征,构建典型相关分析(Canonical correlation analysis,CCA)的Swarm联邦框架,以提取Swarm节点的低维关联特征.为保护协作参数的隐私安全,还构建一种随机扰乱策略来隐藏Swarm特征隐私.在真实数据集上的实验验证了所提方法的有效性.

Federated learning(FL)has significant advantages in solving the problems of privacy disclosure and data islands faced by artificial intelligence(AI).Previous studies on federated learning do not consider the problems of relevance and high dimensionality of data distributed among different federations.Based on the relevance of federated data,a decentralized federated dimensionality reduction method is proposed.This method draws on the idea of Swarm learning(SL).Based on the separation of coupling features,a Swarm federated framework for canonical correlation analysis(CCA)is constructed to extract the low dimensional correlation features of Swarm nodes.In order to protect the privacy of collaboration parameters,a random disturbance strategy is also constructed to hide the privacy of Swarm features.Experiments on real data sets verify the effectiveness of the proposed method.