面向联邦学习的本地差分隐私设计

Towards-federated-learning local differential privacy design

为探究联邦学习的权重聚合框架下本地差分隐私对模型精度影响并提升服务端聚合速度以及本地模型的收敛速度。本文面向联邦学习的本地差分隐私机制的实现探讨cells的个数、ε、客户端数量、扰动机制和数据分配方式对模型精度的影响。本文还设计自适应隐私预算策略,提升模型的收敛速度,方法是使用相邻轮模型的相似性与初始隐私预算建立反比关系从而自适应调整隐私预算。实验表明,从精度损失来看,随cells个数、ε、客户端数量增加而变小;在同等ε下,RR,OUE的结果近似一致,NonIID比IID精度损失率高;自适应隐私预算策略能够根据相邻轮模型的相似度提升模型收敛速度。

To investigate the impact of local differential privacy on model accuracy and improve the server-side aggregation speed as well as the convergence speed of the local model under the weight aggregation framework of federated learning,this paper explores the impact of the number of cells,ε,the number of clients,the perturbation mechanism and the data distribution method on the model accuracy for the implementation of the local differential privacy mechanism for federated learning.This paper also designs an adaptive privacy budget strategy to improve the convergence speed of the model by using the similarity of neighboring round models to establish an inverse relationship with the initial privacy budget and thus adaptively adjust the privacy budget.Experiments show that in terms of accuracy loss,it becomes smaller as the number of cells,ε,and the number of clients increase;the results of RR,OUE are approximately the same for equalε,and NonIID has a higher accuracy loss rate than IID;the adaptive privacy budget strategy can improve the model convergence speed according to the similarity of the adjacent round models.