A learned adaptive hybrid elliptical cloaking framework for location, trajectory, and mobility-pattern privacy in location-based services

Abstract

Ensuring robust privacy for locations, trajectories, and mobility patterns in Location-Based Services remains a critical challenge, as existing anonymization techniques often fail to protect these three dimensions simultaneously and typically neglect Point of Interest (POI) awareness and population-density adaptation. This paper proposes a Learned Adaptive Hybrid Elliptical Cloaking (LAHEC) framework comprising two novel mechanisms: EllipCloak, an elliptical cloaking method for location and trajectory anonymization guided by POI cluster direction and population density, and FanEllipCloak, a fan-pattern elliptical approach that extends protection to mobility patterns. Unlike traditional circular or annular cloaking, the framework introduces adaptive cloaking zones aligned with user-selected POIs and population-density clusters, thereby optimizing the trade-off between privacy and utility, while its focused POI-directed scanning strategy offsets the added complexity by significantly reducing the total query-range area. At its core, the framework generates dummy trajectories through a learned generative component built on two conditional models, a diffusion model and a sequence variational autoencoder, which produce realistic decoys that preserved elliptical validators accept or reject so that k-anonymity is guaranteed by construction. Experimental results show that this learned core yields dense and compact dummy sets whose dummy-utility leakage is more than an order of magnitude lower than adversarially generated baselines, while the fan-pattern mechanism sharply lowers the probability of revealing a user’s mobility pattern and further reduces the identification probability as the fan widens. The reported gains are supported by confidence intervals and significance testing, and a lightweight trajectory-smoothing step brings the variational variant to near-ideal resistance against trajectory-prediction attacks, confirming that the proposed framework protects location, trajectory, and mobility-pattern privacy while preserving high query utility.