Supervised Nonlinear Latent Feature Extraction and Regularized Random Weights Neural Network Modeling for Intrusion Detection System

摘要

Colinearity and latent relation among different input features of net work intrusion detection system (IDS) have to be addressed. The strong non-linearity and uncertain mapping between input features and network intrusion behaviors lead to difficulty to built effective detection model for IDS. In this paper, a new supervised nonlinear latent feature extraction and fast machine learning algorithm based on global optimization strategy is proposed to solve these problems. Specifically, for diminishing colinearity among input variables, kernel partial least squares (KPLS) algorithm is employed to extract nonlinear latent features. Then, regularized random weights neural networks (RRWNN) is utilized to construct the intrusion detection model. To optimize the proposed system, the modeling parameters of KPLS and RRWNN are selected in terms of global optimization. Experiments on KDD99 data show that the proposed approach is effective.