Research on Network Security Situation Prediction Based on Improved GAN

Abstract

This paper presents a model of situation prediction using generative adversarial network to simulate the situation development process. In order to solve the problem of GAN (Generative Adversarial Network) convergence difficulty, this paper takes the convergence distance sum function of Wasserstein distance as the theoretical basis for calculating GAN loss function, and the gradient penalty method is also used. Firstly, the simulation dataset to be used will be pre-processed to reduce the over-fitting of the training model due to the mismatch of feature types and the large difference of feature values. Then, a penalty item will be added every time the parameter is updated, aiming to punish the behaviour greater than 1 when the gradient is updated, which can solve the problem of gradient disappearance to some extent. After the improvement of the traditional GAN algorithm, the prediction accuracy and prediction accuracy of the situation value can be greatly improved, which proves the stability of the improved GAN network. Experimental results show that the proposed GAN model has significant advantages over the unimproved GAN model in terms of convergence and prediction accuracy.