Optimal Network Selection Method Using Federated Learning to Achieve Large-Scale Learning While Preserving Privacy
Koki Horita, Bin Yang, Thomas Carette, Masanobu Jimbo, Akihiro Nakao · 2022
These days, smartphones are equipped with multiple radio access technologies (RAT) such as wireless LAN (WLAN) and 4G/5G cellular with autonomous switching between them. However, it is often the case with RAT switching algorithms being based on availability but not on predicted quality of communications, which poses significant frustrations on the part of users, especially when their smartphones connect to poor-quality WLAN automatically. In order to provide uninterrupted communication to smartphone users, it is necessary to predict WLAN quality before switching networks and change the network based on the result. Although machine learning approaches are useful for predicting network quality, conventional machine learning is infeasible to train on large amounts of data collected from users. It is because WLAN data contain private information such as user location information, and it is possible to track user behavior from the training data. In addition, employing complex machine learning algorithms is not possible with the computational resources of a smartphone due to constraints on training and inference time. To the best of our knowledge, no paper implements and evaluates a solution to this problem. To solve this issue, this paper utilizes latent class regression model trained by Federated Learning. Our proposal achieves on-device training on a smartphone with limited computer resource and protects users’ private information. It predicts the quality of access points with various characteristics with high accuracy (95.0% precision, 23.7% recall). Moreover, by developing and installing the application on an Android smartphone, we show that the proposed method reduces the actual application communication disruption time to within 3 seconds at the 75th percentile (85% reduction compared to no functionality) and completes training within 1 second per session on smartphones with limited computational resources.