A Random Access Scheme for Federated Learning Over Massive MIMO Systems
A Random Access Scheme for Federated Learning Over Massive MIMO Systems
复制标题
DOI：
10.1109/jiot.2023.3278256
复制
发表时间：
2023-11
影响因子：
10.6
通讯作者：
Huimei Han;Jun Zhao;Xinyu Zhou
中科院分区：
文献类型：
--
作者：
Huimei Han;Jun Zhao;Xinyu Zhou
针对大规模多输入多输出(MIMO)系统中一些本地设备无法计算其本地模型的问题，提出了一种适用于大规模多输入多输出(MIMO)系统的联合学习(FL)的随机访问(RA)方案。该方案采用多信道模型，允许设备随机选择其上传的信道，然后基站(BS)基于空中计算直接聚合从信道接收的本地模型。我们将这种方案称为大规模MIMO上基于RA的FL(RAFL-MIMO)。此外，为了让更多的设备参与FL过程，我们提出了利用访问类阻止(ACB)方法来选择上传设备，并建立了ACB因子的优化问题。我们还推导了所提出的RAFL-MIMO方案的期望渐近收敛速度，分析表明所提出的RAFL-MIMO方案可以改善FL的性能。基于线性回归、MNIST手写体数字识别、CIFAR-10照片分类的仿真结果表明，所提出的RAFL-MIMO方案明显优于不考虑ACB因子的RAFL-MIMO方案。
In this article, we present a random access (RA) scheme for federated learning (FL) over massive multiple-input–multiple-output (MIMO) systems to tackle the issue of some local devices not being able to compute their local models. This scheme adopts a multichannel model and allows devices to randomly select their uploading channels, and then the base station (BS) aggregates the local models received from channels directly based on the over-the-air computation. We call this scheme as RA-based FL over massive MIMO (RAFL-MIMO). Furthermore, to enable more devices to be involved in the FL process, we propose to utilize an access class barring (ACB) method to select the uploading devices and formulate an optimization problem of the ACB factor. We also derive the expected asymptotic convergence rate of the proposed RAFL-MIMO scheme to analytically show that the proposed RAFL-MIMO scheme can improve the performance of FL. Simulation results based on ${L2}$ -norm linear regression, and MNIST handwritten digits identification, Cifar-10 photograph classification show that the proposed RAFL-MIMO scheme significantly outperforms the case of the RAFL-MIMO without the ACB factor.