Исследование подходов к синтезу и детектированию клонированных голосов (DeepFake)

Abstract

Modern methods of protecting personal information often uses the voice biometric data of the owner of the information to identify the user. When the owner of the information voices the passphrase, he confirms his identity. However, attackers take advantage of the imperfection of such systems and develop methods for voice cloning, to create a twinkly voice for a cyberattack on personal data protection systems. Within the framework of this article, an attempt is made to explore existing methods for detecting cloned voices in order to protect information and counteract cyberattacks. Also, to achieve results, detection systems will be tested on a sample of Russian-language voice recordings taken from open sources. A comparative assessment of existing approaches is carried out in terms of their practical applicability. In particular, the requirements for the occupied memory of a computing device, computational complexity, complexity in implementation and data collection for training were taken into account. In addition, an analysis of the existing prerequisites and trends for the use of voice synthesis and substitution systems was carried out, potential risks were described, and examples of possible damage from the theft of biometric data were given. An attempt was also made to describe the experimental procedure for evaluating the performance of the considered methods with specifying and clarifying conditions. The criteria for verification and validation of the results are set, which allow drawing conclusions about the efficiency of the systems.