A Network Security Situation Prediction Model Enhanced by Multi Head Attention Mechanism

Abstract

The Internet has grown as a result of information technology advancements, and cybercrime is becoming more and more common. To improve the network defense against all kinds of network attacks and reduce the success rate of cyber crimes, the research innovatively proposes to use the multi-tease attention mechanism to improve the gating cycle unit, and use the multi-head attention mechanism to obtain network security feature information at different locations, so as to improve the learning characteristics of network situation prediction and realize network security situation prediction. Three layers comprised the model: the prediction layer, the transform layer, and the circular network layer. The circular network layer was responsible for dimensionality reduction of network information data. Information features were extracted via the transform layer. The outcomes of the predictions were output by the prediction layer. The study's model performed better when taught in both directions, according to the data, and its accuracy could reach roughly 93.5%. The highest level of model accuracy could be reached when other parameters were fixed and the neurons in the feed-forward layer was 28. Compared with other network security situation prediction models, the proposed model could improve the prediction accuracy to around 93.5% and the precision to around 91% on the UNSW-NB15 dataset, while maintaining the F1 value of the model at around 92%. The research-designed model can accurately predict the network security situation changes, which improves the Internet's defense against attacks and maintains the normal operation of the Internet community.