A Network Data Reinforcement Method Based on the Multiclass Variational Autoencoder

Abstract

Anomaly detection models based on deep learning come up against difficulties on the deployment in real scenarios such as generalization problem. The performance of the model based on specific dataset is not as good as expected in other scenarios. In order to avoid this problem, it is a feasible solution to collect network data from the target environment to train the model. This paper proposes a network data reinforcement method based on the multiclass variational autoencoder to complete training tasks with little amount data. In this paper, anomaly detection models based on MLP and CNN are designed, respectively, and validation experiments are carried out on the CICIDS-2018 dataset. Compared with unreinforced models, models based on this method get faster convergence speed during training. During evaluation, models based on this method achieve an average accuracy of 93.69%, while unreinforced models only get an average accuracy of 55.63%. In addition, this method provides competitive results on insufficient data compared with those existing models on sufficient data.