A correlation method for handling infrequent data in keystroke biometric systems

Many applications need methods for handling missing or insufficient data. This paper applies a correlation technique to improve the fallback methods previously used to handle the paucity of keystroke data from the infrequently used keys in a keystroke biometric system. The proposed statistical fallback model uses a correlation based fallback table based on the linear correlation between pairs of keys. Two large long-text keystroke databases are used in the study - one to construct the model and the other to evaluate system performance as a function of sample length.

1. Introduction

A number of applications need methods for handling missing or insufficient data. In speech and language processing, several methods of handling missing or sparse data are described in [2]. The -gram model of missing or infrequent data is estimated based on the -gram model of sufficient data recursively in the back-off [3] and deleted interpolation [4]. Although both models fail if the unigram is missing, this occurs rarely.