With the continuous development of the network environment, the traffic data in the network increasingly presents high-dimensional, huge and complex characteristics, and the network threat is also increasing, the network information security threat prediction and defense mechanism plays an irreplaceable position in network security. Based on the general process of network anomaly detection, combined with deep learning algorithms, the article proposes a network anomaly detection method based on data enhancement to improve the detection accuracy of network anomaly detection model. Self-attention mechanism is embedded in the neural network framework to accomplish the improved SA-GRU network information security threat prediction method. In the performance index comparison experiments of network security posture values predicted using different prediction models, the average absolute error of the training data of the results predicted by this paper’s model is 0.00266, and the average absolute error of the test data is 0.00369, and the prediction accuracy of this paper’s model prediction is significantly higher than that of other deep learning methods. This verifies the effectiveness of the method proposed in this paper. Finally, based on the experimental results, the network information security defense mechanism is proposed from the three levels of data encryption, the use of secret keys and intrusion detection.