Combinação autoajstada de modelos de aprendizagem de máquina com otimização de hiperparâmetros para detecção de intrusos em redes de computadores

Abstract

The increase in security incidents reported in recent years accompanies the expansion of information systems in institutional/corporate environments and the growth in the use of the Internet. In response to this situation, various initiatives are concerned with offering protection to digital environments, seeking to avoid the occurrence of intrusions in computer networks. Within the scope of these initiatives, the work proposed in this thesis consists of a generic method that employs the autonomous/self-adjusting combination of machine learning solutions for the detection of intrusions in computer networks. The proposed generic method makes use of hyperparameter optimization applied to machine learning modeling and is validated by the implementation of an instance of it composed of a Multi-layer Perceptron network and the K-Nearest Neighbors method; the Multi-layer Perceptron network, after being the target of a hyperparameter optimization process, has its classification carried out in conjunction with the K-Nearest Neighbors method according to autonomous criteria for composing the classification of both models. Through experiments, it was concluded that the proposed solution presents im provements in terms of performance metrics in relation to models of combination of classifiers aimed at intrusion detection in computer networks, with results comparable to other solutions proposed in recent literature.