On the security of a neural network-based biometric authentication scheme

We show that neuron weights used in neural network-based biometric authentication scheme defined in GOST R 52633 standard series contain all the information on biometric data and secret key of the legitimate user. So, the complexity of evaluating (with known tables of neuron weights) the legitimate user's secret key is equivalent to the complexity of evaluating one solution of a corresponding system of linear inequalities. Thus, first, neuron weights should be considered as a part of a secret key of the authentication system, and, second, several methods for neural networks protection proposed in the standard are inefficient.