基于深度学习的语义编码虚假轨迹重构方法OA

Background knowledge attacks pose serious threats to trajectory privacy by exploiting prior knowledge to infer user behavior patterns.Existing trajectory reconstruction methods,however,suffer from two major limitations.First,they fail to address semantic location information leakage effectively,and their deep learning models lack opti-mization under noisy conditions,leading to inadequate reconstruction accuracy and weak semantic extraction.Second,their poor generalization ability hinders efficient reconstruction across heterogeneous datasets,thereby limiting the comprehensiveness of privacy protection.To address these limitations,this study proposes a deep learning-based se-mantic encoding method for synthetic trajectory reconstruction(DL-SESTR),a false trajectory reconstruction method based on semantic information encoding.The method integrates a bidirectional long short-term memory network(Bi-LSTM)with an attention mechanism to capture spatiotemporal dependencies and dynamically identify key trajectory points,thereby improving noise resistance.It also introduces a point-of-interest semantic annotation algorithm(PSA)that matches multi-source point-of-interest(POI)data efficiently to enhance annotation performance.Furthermore,a hierarchical semantic encoding algorithm based on the Hasse diagram(HDSE)is proposed,constructing a semantic sensitivity weight model to distinguish high-priority semantic information from noise.Experiments on the T-Drive and GeoLife datasets evaluated model performance across dense and sparse regions,varying privacy budgets,and day-and-night scenarios.DL-SESTR consistently outperforms baseline methods in balancing privacy protection and data utility:Hausdorff distance is reduced by 0.3％,dynamic time warping(DTW)efficiency improves by 1.2 times,and root mean square(RMS)improves by 1.18 times.Under a low privacy budget(ε=0.01),the method still achieves a 95％Euclidean distance reduction rate,demonstrating strong robustness and generalization ability.