Detalhes bibliográficos
Ano de defesa: |
2018 |
Autor(a) principal: |
LEITE, Urbanno Pereira de Siqueira
 |
Orientador(a): |
FERREIRA, Tiago Alessandro Espínola |
Banca de defesa: |
CAVALCANTI, George Darmiton da Cunha,
GARROZI, Cícero,
MIRANDA, Péricles Barbosa Cunha de |
Tipo de documento: |
Dissertação
|
Tipo de acesso: |
Acesso aberto |
Idioma: |
por |
Instituição de defesa: |
Universidade Federal Rural de Pernambuco
|
Programa de Pós-Graduação: |
Programa de Pós-Graduação em Informática Aplicada
|
Departamento: |
Departamento de Estatística e Informática
|
País: |
Brasil
|
Palavras-chave em Português: |
|
Área do conhecimento CNPq: |
|
Link de acesso: |
http://www.tede2.ufrpe.br:8080/tede2/handle/tede2/7861
|
Resumo: |
In this paper, feature selection models are developed for keystroke dynamics on mobile devices. Two models are elaborated, one based on Genetic Algorithm (GA) and another on PSO. The proposed selectors are applied to a public database of keystroke dynamics built from mobile devices. Both methods of selection are used in conjunction with various classifiers: Naive Bayes, Bayes Net, C4.5 (decision tree), Random Forest, k-NN, Support Vector Machine (SVM) and Multilayer Perceptron (MLP). The feature selection methods developed here are evaluated by accuracy, false positive rate (FAR), false negative rate (FRR) - all these measures are obtained from the classifications - and also by the reduction rates of characteristics. The results obtained from the execution of several experiments show that the proposed models were able to add improvements to the measures of performance - when compared to the results of the classifications without selection -, besides reaching high levels of reduction of characteristics. Through a comparative analysis it was also possible to verify that the models developed in this work have performances compatible with other selectors already available in the literature. The proposed methods also call attention for the stability of their behavior, in such a way that the results generated by them have low indices of variability. In this work it was possible to identify the most selected features and also those less chosen by the models, showing that an attribute can be quite selected for a particular classification method, but not so chosen for another classifier. Already analyzing the frequency of selection of characteristics according to their type, it was verified that the two characteristics most selected by the proposed selection methods are attributes inherent to mobile devices. |