An investigation of the predictability of the Brazilian three-modal hand-based behavioural biometric: a feature selection and feature-fusion approach

Abstract

New security systems, methods or techniques need to have their performance evaluated in conditions that closely resemble a real-life situation. The effectiveness with which individual identity can be predicted in different scenarios can benefit from seeking a broad base of identity evidence. Many approaches to the implementation of biometric-based identification systems are possible, and different configurations are likely to generate significantly different operational characteristics. The choice of implementational structure is, therefore, very dependent on the performance criteria, which is most important in any particular task scenario. The issue of improving performance can be addressed in many ways, but system configurations based on integrating different information sources are widely adopted in order to achieve this. Thus, understanding how each data information can influence performance is very important. The use of similar modalities may imply that we can use the same features. However, there is no indication that very similar (such as keyboard and touch keystroke dynamics, for example) basic biometrics will perform well using the same set of features. In this paper, we will evaluate the merits of using a three-modal hand-based biometric database for user prediction focusing on feature selection as the main investigation point. To the best of our knowledge, this is the first thought-out analysis of a database with three modalities that were collected from the same users, containing keyboard keystroke, touch keystroke and handwritten signature. First, we will investigate how the keystroke modalities perform, and then, we will add the signature in order to understand if there is any improvement in the results. We have used a wide range of techniques for feature selection that includes filters and wrappers (genetic algorithms), and we have validated our findings using a clustering technique.