Evaluation of signature verification reliability based on artificial neural networks, Bayesian multivariate functional and quadratic forms

An experimental comparison of various functional neural networks for signature verification is performed. A signature database for the realization of the computing experiment is built. It is confirmed that up to a certain point, the increase of the decision rule dimension reduces the probability of signature verification error, with an increase in the number of neurons in the network reducing the number of errors. A higher-dimension multi-dimensional Bayes functional with stronger inter-feature correlation is found to perform better. The best result for the signature verification is obtained using networks of Bayesian multidimensional functional, with false acceptance rate of $FRR = 0.0288$ and false rejection rate of $FAR = 0.0232$.