At present, many terminal devices or users adopt an access control mode of “one-time identity authentication and access rights assignment, valid for a long time ”. There are risks such as legal equipment or users being impersonated by a malicious user, and legitimate users launching internal malicious attacks. In this paper, we construct a dynamic access control mechanism based on short-term token and user trust. Firstly, design a short-lived access token that is generated based on device attributes and authentication information. Secondly, a user behavior analysis model based on deep convolutional neural networks is established. The user's trustworthiness judgment is realized through the analysis of user access behavior data, and the token update method is set up in combination with the user trustworthiness level, used to achieve dynamic access authorization. Experimental verification shows that the mechanism is dynamic and more secure in terms of access control.