Abstract: To address the problems of inadequate feature extraction, poor generalization ability and neglect of correlation between non-adjacent features in existing IoT network device anomaly detection, we propose an unsupervised network traffic anomaly detection method combining adversarial self-encoder and feature clustering. We used graph attention network and gated temporal convolutional network for spatiotemporal feature extraction, and proposed a multi-stage layer-by-layer propagation mechanism to enhance the feature extraction of the original data by the model. The model's false alarm rate was effectively reduced by using adversarial exercises on the self-encoders to amplify the anomaly scores, and the K-means algorithm was used for feature clustering. Extensive experiments were conducted on four datasets in this paper to verify the effectiveness of the proposed method.