Highly reliable authentication based on handwritten passwords using hybrid neural networks with protection of biometric templates from being compromised

Abstract

Introduction: Biometrics-to-code converters based on neural networks are the ideological basis for a series of GOST R 52633 standards (unparalleled anywhere in the world) and can be used in the development of highly reliable biometric authentication and electronic signature with biometric activation. Purpose: Developing a model of a biometrics-to-code converter for highly reliable biometric authentication by handwritten passwords with high resistance to attacks on knowledge extraction. Results: We demonstrated the vulnerability of neural networks which makes it possible to perform quick directed enumeration of competing examples in order to compromise a biometric pattern and the personal key of its owner. We described a method of effective protection against this attack, and proposed a hybrid model for a biometrics-to-code converter based on a new type of hybrid neural networks, which does not compromise the biometric standard and the user’s key (password), being resistant to such attacks. The high reliability and effectiveness of the proposed model has been experimentally confirmed in handwritten password verification. The reliability indicators for generating a key from a handwritten password were: FRR = 11.5%, FAR = 0.0009% with a key length of 1024 bits (taking into account the presented fakes of a handwritten pattern). Practical relevance: The results can be used in information security applications or electronic document management.