Authentication based on voice passwords with the biometric template protection using correlation neurons
Keywords:
Abstract
Introduction: The issue of protecting biometric data from compromise is closely related to performance issues. Existing methods of biometric voice authentication either do not protect voice data from compromise or give a high percentage of erroneous decisions; in addition, they do not provide resistance to voice image drift. Purpose: To develop a method of biometric voice authentication that is resistant to the drift of biometric data while ensuring the confidentiality of voice parameters. Results: We propose an authentication method using neural network “biometrics-to-code” converters based on a modified model of correlation neurons and their training algorithms. It has been established that correlations between features contain information about images that does not duplicate the information contained in the features. The biometrics-to-code converter based on correlation neurons produces a much lower percentage of errors and several times longer key length than the classical model based on the GOST R 52633.5 learning algorithm. The number of errors was: 3.26%. When the subject's state changes (intoxication or sleepiness), the number of errors for the developed method does not increase as significantly as for the classical model of the biometrics-code neural network converter. Practical relevance: The results can be used to increase the security of computer resources from unauthorized access and biometric data from compromise. Discussion: Combining neurons of various types into a single layer will make it possible to create more stable and reliable biometric-to-code neural network converters.
