Mixed Wavelet-Based Neural Network Model for Cyber Security Situation Prediction Using MODWT and Hurst Exponent Analysis

Previous models have achieved some breakthroughs in cyber security situation prediction. However, improving the accuracy of prediction, especially long-term prediction, is still a certain challenge. Maximal Overlap Discrete Wavelet Transform (MODWT) with strong ability of information extraction can capture the correlation of the time-series better. Mixed Wavelet Neural Network (WNN) architecture with both Morlet wavelets and Mexican hat wavelets can provide excellent localization and scale detection simultaneously. In this paper, MODWT method and mixed WNN architecture are combined to develop a WNN-M prediction model through data-driven approach. In addition, Hurst exponent is utilized to analyze the predictability of decomposed components for removing poor components. To demonstrate the effectiveness of proposed WNN-M model, 12-hour prediction is considered in a real attack scenario named DARPA given by MIT Lincoln Lab. Experimental results show that the R-2 of WNN-M can be improved by 19.87% and the RMSE of WNN-M can be reduced by 4.05% compared with that of traditional WNN model.