A Novel Personal Entropy Measure confronted with Online Signature Verification Systems´ Performance

In this paper, we study the relationship between a novel personal entropy measure for online signatures and the performance of several state-of-the-art classifiers. The entropy measure is based on local density estimation by a hidden Markov model. We show that there is a clear relationship between such entropy measure of a person´s signature and the behavior of the classifier. We carry out this study on a dynamic time warping classifier, a Gaussian mixture model and a hidden Markov model as well. It is worth noticing that the HMM classifier differs from the HMM used for entropy computation. Signatures were split into three categories according to their entropy value. These categories are coherent across four different databases of around 100 persons each: BIOMET, MCYT-100, BioSecure data subsets DS2 and DS3. We studied the impact of such categories on classifier´s performance with a larger signature data subset of DS3, of 430 persons.