Due to its wide-area and high-precision advantages, Global Navigation Satellite System (GNSS) timing is widely employed in critical infrastructures such as power, communication and transportation, maintaining high-precision time synchronization for the system. Nevertheless, due to the lack of authentication and unencrypted structure of civilian GNSS signals, GNSS receiver is vulnerable to be attacked, resulting in disastrous consequences. Therefore, detecting and mitigating a time synchronization attack (TSA) to improve the security of GNSS timing and ensure the normal operation of critical infrastructures is of great significance. We proposed a TSA detection and mitigation algorithm based on long short-term memory (LSTM) neural network. Based on the good nonlinear mapping ability and high self-learning ability of LSTM, the authentic trend of the receiver clock can be learned and clock state can be predicted. Based on the difference between the predicted and measured clock state of the receiver, TSA detection and mitigation can be realized. Experiments and results show that the proposed algorithm can detect and mitigate two well-known types of TSA. In Type I TSA case, the root-mean-square error (RMSE) is improved by 56.41, 89.14 and 0.01 compared with Robust Estimator (RE), Time Synchronization Attack Rejection and Mitigation (TSARM) method and Multi-Layer Perceptron (MLP) neural network, respectively. In Type II TSA case, the RMSE is improved by 41.80, 88.16 and 0.33 compared with RE, TSARM and MLP, respectively. The research results can be applied to time synchronization systems of critical infrastructures, which can improve time synchronization accuracy and security performance.

全球导航卫星系统（GNSS）授时以其广域、高精度的优势，广泛应用于电力、通信、交通等关键基础设施，为系统保持高精度的时间同步。然而，由于民用GNSS信号缺乏认证且结构未加密，GNSS接收器很容易受到攻击，造成灾难性的后果。因此，检测并缓解时间同步攻击（TSA）对于提高GNSS授时安全性、保障关键基础设施的正常运行具有重要意义。我们提出了一种基于长短期记忆（LSTM）神经网络的 TSA 检测和缓解算法。基于LSTM良好的非线性映射能力和较高的自学习能力，可以学习接收机时钟的真实趋势并预测时钟状态。根据接收器的预测时钟状态与测量时钟状态之间的差异，可以实现 TSA 检测和缓解。实验和结果表明，所提出的算法可以检测和缓解两种众所周知的 TSA 类型。在Type I TSA案例中，与鲁棒估计器（RE）、时间同步攻击拒绝和缓解（TSARM）方法以及多层感知器（MLP）方法相比，均方根误差（RMSE）分别提高了56.41、89.14和0.01分别是神经网络。在Type II TSA情况下，与RE、TSARM和MLP相比，RMSE分别提高了41.80、88.16和0.33。研究成果可应用于关键基础设施的时间同步系统，提高时间同步精度和安全性能。