面向多租户数据中心的联邦学习架构下通信开销优化方法

Optimization method for communication overhead in federated learningarchitecture for multi-tenant data center

为降低多租户数据中心联邦学习架构下的高通信开销问题,提出一种基于三元演化模型参数的通信开销优化算法。首先,建立面向多租户数据中心的联邦学习架构模型来实现数据隐私保护;其次,针对联邦学习架构的引入导致租户和数据中心交互产生了过高的通信开销问题,提出一种基于三元演化模型参数的通信开销优化算法,通过结合最优局部模型和三元向量化模型参数的演化方向来减少租户与数据中心模型参数传输之间的冗余通信;同时,基于联邦学习的隐私研究论证分析了在传输通信过程中所提算法能有效保障参与训练租户的隐私信息。最后,实验结果表明,所提方法在保障训练精度的前提下,相比于联邦平均对比算法能有效降低30%的冗余通信开销。

To address the issue of high communication cost under the federated learning framework in multi-tenant data centers,this paper proposed an optimization algorithm based on ternary evolutionary model parameters.Firstly,it constructed a federated learning architecture tailored to multi-tenant data centers for data privacy protection.Secondly,in response to the excessive communication overhead stemming from the implementation of the federated learning framework,which increasing interactions between tenants and the data center,it proposed an optimization algorithm that utilized ternary evolutionary model parameters.This algorithm aimed to reduce redundant communication in the exchange of model parameters between tenants and the data center by integrating the optimal local model with the evolutionary direction of ternary vectorized model parameters.Moreover,by analyzing privacy research based on federated learning,the algorithm effectively ensured the privacy of tenants participating in the training during the communication process.Finally,experimental results demonstrate that,while maintaining training accuracy,the proposed method can effectively reduce redundant communication costs by 30%compared to the federated averaging baseline algorithm.