Reputation-enabled Federated Learning Model Aggregation in Mobile Platforms
Reputation-enabled Federated Learning Model Aggregation in Mobile Platforms
复制标题
DOI：
10.1109/icc42927.2021.9500928
复制
发表时间：
2021-06
期刊：
影响因子：
--
通讯作者：
Yuwei Wang;B. Kantarci
中科院分区：
文献类型：
--
作者：
Yuwei Wang;B. Kantarci
联合学习（FL）建立在参与节点的移动的网络上，这些节点训练本地模型并在中央服务器上为学习模型参数做出贡献，而不必共享其原始数据。服务器聚合上传的模型参数以生成全局模型。上传的本地模型的常见做法是均匀加权聚合，假设网络的每个节点对推进全局模型的贡献相等。由于设备和收集的数据的异质性，用户对全局模型的贡献之间不可避免地存在差异。因此，用户（即，具有较高贡献的设备）在聚集期间应当被赋予较高的权重。考虑到这一点，本文提出了一种声誉使能的聚合方法，规模的聚合权重的用户的声誉分数。用户的信誉分数是根据每个训练轮期间其训练的本地模型的性能度量来计算的，因此它可以是评估其训练的本地模型的直接贡献的度量。建议的聚合方法的基线，利用标准的平均值，以及第二个基线的范围是一个基于声誉的客户端选择的数值比较，显示了17.175%的改进，超过标准基线的非独立和同分布（非IID）的情况下，FL网络的100名参与者。还示出了在用户范围从20到100的较小FL网络下的第一和第二基线的一致改进。
Federated Learning (FL) builds on a mobile network of participating nodes that train local models and contribute to the learning model parameters at a central server without being obliged to share their raw data. The server aggregates the uploaded model parameters to generate a global model. Common practice for the uploaded local models is an evenly weighted aggregation, assuming that each node of the network contributes to advancing the global model equally. Due to the heterogeneous nature of the devices and collected data, it is inevitable to have variations between the contributions of the users to the global model. Therefore, users (i.e., devices) with higher contributions should be weighted higher during aggregation. With this in mind, this paper proposes a reputation-enabled aggregation methodology that scales the aggregation weights of users by their reputation scores. Reputation score of a user is computed according to the performance metrics of their trained local models during each training round, therefore it can be a metric to evaluate the direct contributions of their trained local model. Numerical comparison of the proposed aggregation methodology to a baseline that utilizes standard averaging as well as a second baseline that is scoped to a reputation-based client selection shows an improvement of 17.175% over the standard baseline for not independent and identically distributed (non-IID) scenarios for an FL network of 100 participants. Consistent improvements over the first and second baselines under smaller FL networks with users ranging from 20 to 100 are also shown.