Presentation Information
[BPO-1-05]Implementation and Evaluation of a Hierarchical Federated Learning System for Base-Station Assignment Prediction on Android Devices
◎△Tetsuya ITO Ito1, Norihiko Shinomiya1 (1. Soka Univ.)
Keywords:
Federated Learning,Hierarchical Federated Learning,Base Station Assignment,Android
Federated Learning (FL) has been attracting attention for its potential application to wireless access control. However, most existing studies rely on simulations, and empirical evaluations on real mobile devices remain scarce. In this paper, we report the results of operating a three-tier Hierarchical Federated Learning (HFL) testbed consisting of one central server, two edge servers, and four Android devices on a base station (AP) assignment prediction task. Aggregating results from nine synchronized sessions across four devices, the overall average accuracy gradually improved from 0.649 at cycle 1 to 0.678 at cycle 5, confirming a learning effect through HFL aggregation. However, a device-level analysis revealed asymmetric behavior: devices with initially low accuracy improved by +0.10 to +0.17, while devices with initially high accuracy degraded by approximately −0.08. This heterogeneity in learning effects across devices, which is not observed in simulations, is attributed to non-IID local data distributions, the aggregation characteristics of FedAvg, and the effects of OS scheduling and communication delays inherent to real device environments.