A machine learning approach using nonlinear ARX neural networks with Bayesian regularization for epidemic malware dynamics in critical network infrastructures

A machine learning approach using nonlinear ARX neural networks with Bayesian regularization for epidemic malware dynamics in critical network infrastructures

연구 분야: Analysis

학회: International Journal of Information Security

The rapid growth in digitalization is a primary factor of advancement and expansion in malware attack surfaces in critical network infrastructures. Consequently, in the present era, modelling of these modern trends in malware propagation have utmost interest for the research scholars to mitigate its adverse impact on industrial, strategic and commercial sectors. In this study, an innovative machine learning driven neuroarchitecture is developed for modelling the dynamics of malware propagation in critical network architectures by integrating the nonlinear multilayer autoregressive exogenous neural networks (NARXNN) with Bayesian regularization (BR) i.e. NARXNN-BR algorithm. The proposed NARXNN-BR technique is implemented on an epidemic nonlinear malware propagation (ENMP) model constructed on seven dynamic states: susceptible, delitescent, infected, quarantine, traced, patched and recovered nodes to investigates the chronological dependencies and complicated interactions of malware spread. The synthetic data for ENMP model is generated via Adams numerical solver to analyze the dynamics of malware spread in the networks corresponding to the distinct case studies such as variation in the rate of suspectable to infected nodes, the rate of delitescent to infected nodes, the rate of recovered to susceptible nodes, the rate of susceptible to recovered nodes, the rate of delitescent to isolated nodes and the rate of traced to patched nodes. The machine learning predictive NARXNN-BR system is executed for diverse case studies of ENMP model by randomly perturb data for training, testing and validation sets to formulate a solution network through minimization of MSE in the range 10–09 to 10–10. The robustness of presented NARXNN-BR methodology is substantiated by comparative evaluation on convergence tendencies of MSE metric, correlation assessment between inputs and outputs, histograms analysis of error and autocorrelation indices for error to investigate the ENMP system.