The reduction of size requirements of test biometric samples while transition to using the Bayes multidimensional correlation functionals

The problem of improving the reliability of biometric authentication procedures of users of computer systems using small training sample for configuring of recognition machines was considered. In the present article is shown that the Bayes multidimensional correlation functionals can be modified to using the identically correlated in modulus biometric data. The correlation coefficients of biometric data have a significant error at a small test samples. This impedes their use when configuring the classical quadratic form and Bayesian networks. The authors suggested using the method of symmetrization of biometric data, ie, replace biometric data by other multidimensional data with the same entropy. After symmetrization of correlation matrix it will have the same coefficients of pair correlations outside the diagonal. It is proved that requirements to biometric data volume are significantly reduced while using the mentioned method, error of calculation results of correlation coefficients gets smaller. The effect of increasing the stability of calculations is observed for any biometric data. As a consequence, the configuring of the classical quadratic forms and of the Bayesian maximum likelihood networks become more resistant tasks. The data that allowing to estimate how much the size of biometric data sample can be reduced are provided.