Privacy-Conscious Location-based Service using Federated Analytics and Secure Computation

Abstract

This project proposes a privacy-preserving framework for location-based services, incorporating federated analytics, secure computation and mobility intelligence for customized next-location prediction and recommendation of points-of-interest. The proposed system is based on the Foursquare NYC check-in data set of user-wise location traces, venue categories, geographic coordinates and timestamps. The whole dataset is split by user identity to simulate federated clients and all user check-in logs are not stored at a central server. Clients train an LSTM-based sequence model to learn temporal mobility behavior from their check-in patterns locally. Gaussian noise is applied using a differential privacy layer prior to the sharing of model updates, to minimize the likelihood of revealing information about the location of the user. The noisy local updates are then aggregated at the Flask server using the FedAvg algorithm, which leads to an aggregated mobility model without noise. In addition to next-location prediction, the system uses matrix factorization to produce top 5 recommendations and DBSCAN clustering to detect spatial hotspots based on spatial patterns. The frontend is created with HTML, CSS, JavaScript and Leaflet.js and presents recommended venues, predicted latitude/longitude, hotspot heatmap and an epsilon-based privacy budget meter. The novelty of the work is that the location prediction, POI recommendation, hotspot detection, federated aggregation, and differential privacy are all integrated into a unified location-based service architecture. This not only facilitates beneficial mobility analytics but also minimises the central exposure of sensitive GPS traces and enhances the privacy awareness of location-based recommendation systems.