The vehicle-to-infrastructure (V2I) communications enable a wide range of new applications, which bring prominent benefits to the transportation system. However, malicious attackers can potentially launch falsified data attacks against V2I applications to jeopardize the traffic operation. To ensure the benefits brought by the V2I applications, it is critical to protect the applications from those cyber-attacks. However, existing literature on the defense solution that protects the V2I applications is very limited. This paper aims to fill this research gap by proposing a data-driven method to identify falsified trajectories generated by compromised connected vehicles (CVs). A trajectory embedding model, inspired by the word embedding model from the natural language processing (NLP) community, is developed. The proposed trajectory embedding model generates vector representations of vehicle trajectories that can be used to compute the similarities between trajectories. The proposed method consists of two steps. In the first step, historical trajectory data are used to train a neural network and obtain the vector representations of trajectories. The second step computes a distance matrix between each pair of trajectories and identifies falsified trajectories using a hierarchical clustering algorithm. Simulation experiments show that the proposed method has a very high detection rate (>97.0%) under different attack goals with varying CV penetration rates from 100% to 25%, while the false alarm rate remains low. It has great potential to be implemented in a wide range of trajectory-based CV applications such as traffic state estimation and traffic signal control, to safeguard the CV system from cyber threats.

车辆到基础设施（V2I）通讯实现了广泛的新应用，这为运输系统带来了显着的好处。但是，恶意攻击者可能会针对V2I应用程序发起伪造数据攻击，从而危害流量操作。为了确保V2I应用程序带来的好处，保护应用程序免受那些网络攻击至关重要。但是，有关保护V2I应用程序的防御解决方案的现有文献非常有限。本文旨在通过提出一种数据驱动的方法来识别由受损的联网车辆（CV）产生的伪造轨迹来填补这一研究空白。开发了一种轨迹嵌入模型，该模型受自然语言处理（NLP）社区的单词嵌入模型的启发。提出的轨迹嵌入模型生成车辆轨迹的矢量表示，可用于计算轨迹之间的相似性。所提出的方法包括两个步骤。第一步，使用历史轨迹数据来训练神经网络并获得轨迹的矢量表示。第二步计算每对轨迹之间的距离矩阵，并使用分层聚类算法识别伪造的轨迹。仿真实验表明，该方法在不同的攻击目标下具有很高的检测率（> 97.0％），CV渗透率从100％到25％不等，而虚警率仍然很低。