Graph Convolutional Networks (GCNs) utilize topology and content information for attribute network community detection. However, there are issues such as insufficient adaptation to unsupervised learning, limited depth of GCN layers, and inadequate information fusion in the task. Therefore, this paper proposes a novel community detection model named AE-GAN (Autoencoder-enhanced Graph Attention Network). Firstly, AEGAN introduces a self-attention mechanism and a self-encoder module to learn node content information, transmitting the acquired content information to the GCN module. Meanwhile, the GCN module integrates the learned content information with the topology information through convolution, extending the depth of the GCN layer and solving the problem of imbalance in the fusion of the two types of information. Finally, we propose a triple self-supervised module under this model to automatically guide unsupervised learning tasks. In a series of comparative experiments, AEGAN demonstrates significant improvements in evaluation metrics such as ACC and NMI, highlighting its effectiveness in attribute network community detection.