基于Fed-DPDOBO的分散式联邦学习

Decentralized Federation Learning Based on Fed-DPDOBO

传统的客户-服务器架构联邦学习作为解决数据孤岛问题的有效手段,其中心服务器面临着巨大的带宽压力,分散式的对等架构联邦学习在一定程度上可改善这种情况。然而,联邦学习的客户端还存在着数据隐私泄露的风险,而且其成本函数梯度信息在某些情况下很难获得。针对这些问题,本文为一致性约束下的对等架构联邦学习设计一种Federated Differential Privacy Distributed One-point Bandit Online(Fed-DPDOBO)算法,可有效地解决中心服务器带宽限制和客户端梯度信息未知的问题。此外,差分隐私技术的运用,可很好地保护各客户端数据隐私。最后,通过利用MINST数据集进行分散式联邦学习实验,验证本文算法的有效性。

The traditional client-server architecture federation learning is an effective means of solving the problem of data silos,where the central server is under enormous bandwidth pressure and the decentralized peer-to-peer architecture federation learn⁃ing improves this situation to some extent.However,clients of federal learning also suffer from the risk of data privacy breaches and the gradient information of their cost function is difficult to obtain in some cases.To address these issues,this paper designs an Federated Differential Privacy Distributed One-point Bandit Online algorithm(Fed-DPDOBO)for peer-to-peer architecture federation learning under consistency constraints,which effectively addresses the problems of bandwidth limitation of the central node and unknown gradient information of the client.In addition,data privacy for each client is well protected due to the use of differential privacy technology.Finally,the effectiveness of this paper′s algorithm is verified by conducting decentralized federa⁃tion learning experiments with the MINST dataset.