Safe Reinforcement Learning-based Indoor 3D Location Privacy Protection

Abstract

Users share their current locations to the location-based service (LBS) server to request LBS in high-rise indoor buildings. However, this also raises the risk of users' location privacy leakage. An untrusted LBS server can infer users' personal private information and then sends spam or fraudulent messages to the user. Existing works mainly focus on protecting two-dimensional (2D) location privacy, ignoring the height dimension of the user's location data, i.e., three-dimensional (3D) geolocation. Besides, employing a high-risk location privacy protection policy may cause severe privacy leakage. In this paper, we study the 3D location privacy protection mechanism by randomly perturbing the user's location data based on 3D geo-indistinguishability. To better balance the privacy protection and the quality of service (QoS), we propose a scheme to adaptively adjust the perturbation policy based on reinforcement learning (RL) with safe exploration. The core idea of safe exploration is to continue evaluating the risk level of each state-action pair to rule out the high-risk state-action pair. Due to the large number of state-action pairs in real-life situations, we further propose a safe deep RL-based 3D location privacy protection scheme (SDRLP), using convolutional neural networks (CNNs) to better capture the system state and estimate the risk level. Simulation results demonstrate that the SDRLP-based scheme increases privacy and reduces QoS loss compared with the SRLP-based scheme.