Anomaly-resilient geofencing and predictive navigation in IoT environments using machine learning and federated learning for metaverse workplaces and smart shopping malls

TL;DR

This paper introduces systems for smart navigation in workplaces and shopping malls using machine learning and privacy-preserving techniques.

Contribution

A novel framework combining geofencing, anomaly detection, and federated learning for immersive navigation in IoT environments.

Findings

Anomaly detection using One-Class SVM achieves 93.5% accuracy in filtering spurious Wi-Fi signatures.

LSTM-based sequence prediction reaches 59% top-1 accuracy in forecasting user destinations.

Federated Learning preserves privacy with only a 2–5% accuracy loss in model training.

Abstract

The rapid convergence of physical and digital environments is redefining user interactions in both professional and retail sectors. While the concept of the Metaverse offers new avenues for immersive remote collaboration, complex physical venues such as shopping malls require intelligent optimization to mitigate navigational inefficiencies and enhance user satisfaction. This research integrates augmented reality (AR), virtual reality (VR), and the Metaverse alongside machine learning (ML) and Federated Learning (FL) to create virtual spaces for workplace meetings in the Meta Workplaces Monitoring System (MetaWMS) and an active navigation application for shopping malls, the Meta Shopping Navigation System (MetaSNS). To ensure data integrity within these IoT environments, anomaly detection is applied prior to geofencing to filter out spurious Wi-Fi network signatures, such as mobile hotspots. Validated against the Aegean Wi-Fi Intrusion Dataset 3 (AWID3), the proposed One-Class SVM gatekeeper achieves a detection accuracy of 93.5%, significantly outperforming KNN (86.6%) and Isolation Forest (67.4%). Geofencing is then used to define virtual perimeters, enabling location-specific AR experiences. Building on previous work in indoor geofence detection, this paper extends the framework to support intelligent navigation using the large-scale Microsoft Research Indoor Location dataset. Sequence Prediction is performed using a Long Short-Term Memory (LSTM) architecture to forecast users’ next likely destinations, achieving prediction accuracies of 59%, 77%, and 83% for top-1, top-3, and top-5 recommendations, respectively. To preserve privacy, Federated Learning (FL) is employed so that only model weights, rather than raw data, are shared with the server, introducing a marginal accuracy loss of 2–5% while ensuring privacy-preserving personalization.

Figures