The resource access behavior in university teaching system has time continuity, role difference and context dependence, which affect the discrimination accuracy and implementation stability of access control. This paper proposes a neural network driven approach for modeling access behavior. Based on the semester log of a university teaching platform, a data set containing 1.26 million access records, 42,380 valid sessions, 18 types of teaching resources and 6-level permission labels was constructed. By jointly encoding the user role, request interval, resource sensitivity level, terminal type and historical path, the session-level behavior representation is generated by combining the gated recursive unit, graph relationship propagation and attention aggregation mechanism, and the permission discrimination and control execution are completed. Experimental results show that the accuracy of access level identification reaches 97.3%, the F1 value reaches 96.1%, the average response delay is 21 ms, and the unauthorized misjudgment rate is reduced to 0.18%. This method can maintain stable control performance in different teaching stages and different request sizes, and support fine-grained, real-time and computable control of teaching resources access in university teaching systems.
