In the Internet era, the propagation of social networks is becoming increasingly frequent, and the various abnormal situations emerging during propagation are overwhelming. Therefore, it is particularly crucial to maintain the normal propagation of social networks and promptly detect abnormal phenomena in social network propagation. Given their exceptional ability to capture spatial information, graph networks have been widely utilized in the field of propagation. This paper proposes a deep autoencoder framework, termed GUNet-AE, which integrates graph convolution and a U-Net structure, and applies it to the detection of anomalies in real-world social networks. The framework is built upon a deep autoencoder and incorporates the widely-used Graph Convolutional Network (GCN) to learn the underlying distributions of graph structures and node attributes. By leveraging the U-Net structure, the framework learns representations at multiple scales, and detects anomalies by constructing the reconstruction loss based on both structural and attribute losses. We compiled a real dataset from trending topics on Sina Weibo, and the performance of the model has been tested on this dataset as well as several publicly available datasets with positive results.