Nowadays, large-scale individual trajectories can be collected by various location-based social network services, which enables us to better understand human mobility patterns. However, the trajectory data usually contain sensitive information of users, raising considerable concerns about the privacy issue. Existing methods for protecting user trajectory data face two major challenges. First, existing methods generally emphasize on data privacy but largely ignore the data utility. Second, most existing work focus on protecting the privacy of users’ check-in locations, which is not sufficient to protect against the trajectory-user linking (TUL) attack that infers a user’s identity based on her/his trajectories. In this paper, we for the first time propose a collaborative adversarial learning model named BPUCAL to effectively resist the TUL attack and preserve the data utility simultaneously. The general idea is to fool the TUL model by adding a small perturbation on the original trajectory data to balance the data utility and privacy. BPUCAL perturbs a few numbers of carefully identified check-ins of a trajectory which are pivotal for a TUL model to infer the identity of a user. Specifically, BPUCAL contains three parts: a perturbation generator, a discriminator, and a TUL model. The generator aims to produce learnable noise and adds it to the original trajectories for obtaining perturbed trajectories. The perturbed trajectories with a minimal changes compared to the original trajectories can deceive both the discriminator and the TUL model. Extensive experiments are conducted over two real-world datasets. The results show the superior performance of our proposal in balancing data privacy and utility on trajectory data by comparison with baselines.

如今，各种基于位置的社交网络服务可以收集大规模的个人轨迹，这使我们能够更好地了解人类的流动模式。然而，轨迹数据通常包含用户的敏感信息，引起了人们对隐私问题的极大关注。现有的保护用户轨迹数据的方法面临两大挑战。首先，现有方法普遍强调数据隐私，但很大程度上忽视了数据的效用。其次，大多数现有工作都集中在保护用户签到位置的隐私，这不足以防范基于轨迹推断用户身份的轨迹用户链接（TUL）攻击。在本文中，我们首次提出了一种名为BPUCAL的协作对抗学习模型，以有效抵御 TUL 攻击，同时保留数据效用。总体思路是通过在原始轨迹数据上添加一个小的扰动来欺骗TUL模型，以平衡数据效用和隐私。BPUCAL 会干扰一些经过仔细识别的轨迹签入，这对于 TUL 模型推断用户身份至关重要。具体来说，BPUCAL包含三部分：扰动发生器、鉴别器和TUL模型。生成器的目的是产生可学习的噪声并将其添加到原始轨迹中以获得扰动轨迹。与原始轨迹相比变化最小的扰动轨迹可以欺骗鉴别器和 TUL 模型。在两个真实世界的数据集上进行了广泛的实验。结果表明，与基线相比，我们的建议在平衡轨迹数据的数据隐私和实用性方面具有优越的性能。