Attention monitoring is an essential task to evaluate the human cognitive status in human-computer interaction. Prior works either employ an inconvenient invasive method or struggle to provide an explainable mechanism between the original human signal and the attention monitoring results. In this paper, we present Gaze2Atten, a dynamics-based cognitive learning approach for monitoring human attention with the non-invasive gaze signal. Gaze2Atten is constructed based on the dynamic system theory, which makes our mechanism explainable in attention modeling and monitoring. The attention-related gaze dynamics model is first learned based on the cognitive dynamics characteristics of humans. Furthermore, we realize an efficient dynamic pattern-matching method to early detect abnormal attention in the human-computer interaction process. To validate our approach, we designed a serious game to carry out a human-computer interaction behavior, and a parallel gaze data collection of subjects, the analysis shows that Gaze2Atten enables efficient and real-time human attention monitoring.