Federated Learning for WiFi Fingerprinting

Abstract

Channel State Information (CSI) based fingerprinting is surfacing as an accurate and robust method of indoor localization. However, the high-dimensional nature of CSI data impedes its adoption in multi access point (AP) systems. To reap the rewards of cooperative localization with privacy and limited system complexity in mind, we propose a federated learning (FL) architecture. Each AP has an individual model and a shared model, where the individual model parameters are unique to each AP and the shared model parameters are communicated to a central server for aggregation. The server averages the models and sends them back to each AP, which use this joint model as a regularization term. To capture the spatio-temporal characteristics of CSI, we propose a convolutional neural network (CNN) as each AP’s individual model and a multi layer perceptron (MLP) as the shared model. Extensive experimental studies verify the superiority of the proposed edge computing approach compared to the exiting methods in the literature. We use commercial off-the-shelf APs collecting CSI data in multiple indoor environments and compare the proposed system to a state-of-the-art deep learning model. Our approach shows significant improvement in the localization accuracy for both individual APs and aggregate predictions.

Topics & Concepts