A deep learning access control model based on user preferences is proposed to address the issue of personal privacy leakage in social networks. Firstly, social users and social data entities are extracted from the social network and used to construct homogeneous and heterogeneous graphs. Secondly, a graph neural network model is designed based on user daily social behavior and daily social data to simulate the dissemination and changes of user social preferences and user personal preferences in the social network. Then, high-order neighbor nodes, hidden neighbor nodes, displayed neighbor nodes, and social data nodes are used to update user nodes to expand the depth and breadth of user preferences. Finally, a multi-layer attention network is used to classify user nodes in the homogeneous graph into two classes: allow access and deny access. The fine-grained access control problem in social networks is transformed into a node classification problem in a graph neural network. The model is validated using a dataset and compared with other methods without losing generality. The model improved accuracy by 2.18% compared to the baseline method GraphSAGE, and improved F1 score by 1.45% compared to the baseline method, verifying the effectiveness of the model.