Methods of countering speech synthesis attacks on voice biometric systems in banking (review article)

The paper considers methods of countering speech synthesis attacks on voice biometric systems in banking. Voice biometrics security is a large-scale problem significantly raised over the past few years. Automatic speaker verification systems (ASV) are vulnerable to various types of spoofing attacks: impersonation, replay attacks, voice conversion, and speech synthesis attacks. Speech synthesis attacks are the most dangerous as the technologies of speech synthesis are developing rapidly (GAN, Unit selection, RNN, etc.). Anti-spoofing approaches can be based on searching for phase and tone frequency anomalies appearing during speech synthesis and on a preliminary knowledge of the acoustic differences of specific speech synthesizers. ASV security remains an unsolved problem, because there is no universal solution that does not depend on the speech synthesis methods used by the attacker. In this paper, we provide the analysis of existing speech synthesis technologies and the most promising attacks detection methods for banking and financial organizations. Identification features should include emotional state and cepstral characteristics of voice. It is necessary to adjust the user's voiceprint regularly. Analyzed signal should not be too smooth and containing unnatural noises or sharp interruptions changes in the signal level. Analysis of speech intelligibility and semantics are also important. Dynamic passwords database should contain words that are difficult to synthesize and pronounce. The proposed approach could be used for design and development of authentication systems for banking and financial organizations resistant to speech synthesis attacks.