부산대학교 도서관

The rapid growth and evolution of indoor location-based services (LBS) increase the risk of location privacy breaches. Additionally, the semantic tag of locations can easily expose users’ sensitive private information. This paper studies the safe reinforcement learning (RL)-based semantic location privacy protection mechanism (LPPM) in three-dimensional (3D) spaces, such as multi-story buildings. This provides a unique opportunity to derive the optimal location privacy protection (LPP) policy with safe exploration in dynamic 3D environments without knowing the accurate attack model. Safe exploration avoids selecting high-risk perturbation policies by evaluating each state-action pair’s risk level. We propose a safe Dueling Double DQN (D3QN)-based 3D LPPM (SDLPPM) to adaptively explore the perturbation policy, avoiding the overestimation of Q-values and reducing the redundant exploration in similar 3D states. 3D geo-indistinguishability is used to randomly perturb users’ locations. Besides, we further develop a safe A3C-based LPPM (SALPPM) to continuously select the perturbation policies to avoid discretization errors of perturbation angles and privacy budgets. This mechanism uses multi-thread technology to interact with the environment independently to accelerate the policy selection in complex 3D LPP scenarios. Simulation results demonstrate that the proposed mechanisms increase privacy and reduce QoS loss compared to the benchmark mechanisms.