Reduced complexity attack characterisation using discriminant functions for the Gaussian distribution

In this paper we describe a reduced complexity attack characterisation technique. A Bayesian framework is constructed, and the underlying distributions are assumed Gaussian. This allows quadratic discriminant functions to be used. This technique has the advantage over previous non-parametric techniques that histograms derived from Monte Carlo simulations are not necessary. Instead, only the mean and covariance matrix are required for each attack. This allows the number of features to the classifier to be increased providing superior classification performance without posing significant memory or computational requirements. We also show that in many cases the improvements in performance due to not having a fixed histogram bin size or issues with histogram sparsity outweigh the disadvantages due to a mismatch between the model and the observed data.