FGLNoRe: Federated Graph Learning on Non-IID RF Fingerprints for Building-Floor Positioning

In recent years, deep learning has revolutionized indoor positioning based on radio frequency fingerprints. Furthermore, federated learning has been introduced to address the challenges of data sharing without compromising location privacy in indoor positioning. In the context of federated learning, there is the inherent heterogeneity among fingerprint data from diverse mobile clients. Moreover, federated learning approaches overlook the underlying spatial correlations among access points (APs). Therefore, we need to improve model aggregation techniques and integrate graph learning techniques into federated learning to effectively learn from heterogeneous and spatially correlated APs. This paper proposes the FGLNoRe algorithm (Federated Graph Learning via Normalized Averaging with Regularization), a federated graph learning method for fingerprinting-based building-floor positioning in large-scale three-dimensional scenarios. The method introduces federated graph learning via normalized averaging with regularization to integrate federated learning with graph convolutional network (GCN). Extensive experiments on a real Wi-Fi fingerprint dataset validate that our method achieves a building-floor positioning classification accuracy of 0.85 in the presence of high-level data heterogeneity.