<p>Global navigation satellite system (GNSS) time service serves as the core support for high-precision time synchronization in critical infrastructures, including power grids and communication networks. However, GNSS timing receivers are inherently vulnerable to time synchronization attacks (TSA), which can induce unexpected time synchronization deviations and further threaten the operational security of the entire system. To solve this problem, this paper proposes a detection and mitigation method for time synchronization attacks. This method detects and recovers clock bias by training two convolutional neural network—long short-term memory neural networks, one of which is used for short-term clock bias prediction, initially detects the occurrence of attacks, and determines the input of long-term clock bias prediction model; The second network is used for long-term prediction of clock bias to determine whether the clock bias is credible, and it is also used as an attack mitigation means to replace the real clock bias with the predicted one. According to the verification results and analysis on the measured data of the physical time service receiver deception, we found that the detection rate of the dual networks joint model was more than 98%, and the detection rate of each TSA scene is more than 19 times higher than that of a single short-term prediction network detection method. The long-term mitigation performance against TSA fundamentally depends on the long-term prediction accuracy of the receiver clock bias. In the normal operation scene without spoofing, the proposed method achieves a root mean square error as low as 9.04&#xa0;ns for 2&#xa0;h ahead clock bias extrapolation, which significantly outperforms other classical clock prediction algorithms. The proposed method can provide theoretical support and engineering solutions for the security protection of high-precision time service system based on GNSS.</p>