Analysis of machine learning methods for computer systems to ensure safety from fraudulent texts

IT Security is an essential condition for functioning of each company whose work is related to the information storage. Various models for detecting fraudulent texts including a support vector machine, neural networks, logistic regression, and a naive Bayes classifier, have been analyzed. It is proposed to increase the efficiency of detection of fraudulent messages by combining classifiers in ensembles. The metaclassifier allows to consider the accuracy values of all analyzers, involving in the work the construction of the weight matrix and the characteristic that determines the minimum accuracy boundary. Based on the developed method, a software module for the classification of fraudulent text messages written in Java using M1 class of the OPENCV open library was created and tested. The general algorithm of the ensemble method is given. An experiment based on logistic regression, a naive Bayesian classifier, a multilayer perceptron, and an ensemble of these classifiers has revealed the maximum efficiency of the naive Bayesian classification algorithm and the prospect of combining classifiers into ensembles. The combined methods (ensembles) improve the results and increase the efficiency of the analysis, in contrast to the work of individual analyzers.