Abstract:

Aiming at the problem of ensuring the effectiveness and mining analysis value of trajectory semantic data while realizing personalized privacy protection of vehicle trajectory data， a vehicle trajectory semantic protection mechanism based on improved Generative Adversarial Network （GAN） was proposed. In this mechanism： firstly， a position sensitivity grading and semantic annotation method based on Hidden Markov Model （HMM） was designed to extract the effective stop points from vehicle trajectories， and then the stop points were divided into different sensitive levels and annotated semantically. Secondly， Long Short-Term Memory （LSTM） network was introduced into the improved GAN to construct the semantic trajectory model based on the dynamic GAN， and the GAN model was used for training to generate high-quality synthetic trajectories. Finally， for the stop points in synthetic trajectories that required further privacy protection， a differential privacy personalized protection algorithm combining the position sensitivity levels was proposed， which assigned privacy budgets to the stop points according to their sensitivity level and correlation between the stop points， and noise was injected by combining with the Laplace mechanism to achieve the privacy protection， so as to maximize the usability of the trajectory data after protection. Experimental results show that compared to the LSTM-TrajGAN model， the proposed mechanism reduces the Mutual Information （MI） value by 27.58% and improves the semantic trajectory similarity by 24.4%. It can be seen that the proposed mechanism protects user privacy effectively while ensuring the usability of semantic trajectory data.

Key words: intelligent transportation system, trajectory privacy protection, semantic annotation, Generative Adversarial Network (GAN), trajectory synthesis, personalized privacy protection

摘要：

针对在实现个性化车辆轨迹数据隐私保护的同时保证轨迹语义数据的有效性和挖掘分析价值的问题，提出一种基于改进生成对抗网络（GAN）的车辆轨迹语义保护机制。在该机制中：首先，设计一种基于隐马尔可夫模型（HMM）的位置敏感分级语义标注方法，用于从车辆轨迹中提取出有效的停留点，并对停留点进行敏感等级划分和语义标注；其次，将长短期记忆（LSTM）网络引入改进的GAN中，构建基于动态GAN的语义轨迹合成模型，利用GAN模型进行训练以生成高质量的合成轨迹；最后，针对合成轨迹中需要进一步隐私保护的停留点，提出一种结合位置敏感等级的差分隐私个性化保护算法，该算法根据停留点的敏感等级和停留点之间的相关性为停留点分配隐私预算，并且结合拉普拉斯机制注入噪声实现隐私保护，最大限度地保证轨迹数据保护后的可用性。实验结果表明，相较于LSTM-TrajGAN模型，所提出的框架互信息（MI）值降低了27.58%，语义轨迹相似度提高了24.4%。可见，所提机制在保证语义轨迹数据可用性的同时有效保护了用户隐私。

关键词: 智能交通系统, 轨迹隐私保护, 语义标注, 生成对抗网络, 轨迹合成, 个性化隐私保护