The methods of hacking into the voice biometrics systems

Playing speaker recordings

Recording the victims voice by using the tape recorder, smartphone or other device and then recreate such a recording during log in is easily detected by voice biometrics fraud system. This type of attack is indeed dangerous – independent studies show a high susceptibility user verification systems for this type of intrusion. However, a lot depends on the quality of the used recording. In the case of VoicePIN, the sound output from the speaker is not suitable to use for logging in – it reduces significantly the risk of fraud. Speaker verification system run mode with fixed passwords allows the use of techniques to counter attacks by replay. Fraud detection mechanism is
based on the premise that human speech expresses the same content each time in a slightly different way. If the log in attempt is detected by means of a recording which is too similar to the sample previously used method in the system, the verification is negative. The used method is highly effective (EER of 0.6%).
The second possibility to protect themselves against this type of attack is to analyze the acoustic characteristics and detection of the use of recordings in the verification process. This method offers efficiency at the level of 5-8% EER.

Speech synthesis

Speech synthesis systems (called Text-to-Speech or TTS) consist of two main modules: text analysis and sound generation. The latter has a direct impact on the sound of synthesized speech and that depends on the efficiency of the system as a tool to attack the voice biometrics. Since the 70s the technology used to generate speech has evolved and the latest system is already so advanced that it can well imitate a particular speaker. These systems, thanks to models based on the theory of processes classified by Markov (sequences of events, in which the probability of each event depends only on the result of a previous event) can learn how to speak like a human using a relatively small amount of training recordings. Tests referred in the literature have shown that this type of attack is highly dangerous (FAR increase from 0 to 81%).

However there are methods of detecting synthesized speech. Natural voice is characterized by volatility dynamics that synthetic speech are considerably smaller. The differences between natural
and synthesized speech are noticeable in high frequencies, which mentions natural is more volatile and therefore the characteristics exhibit a higher variance. The fundamental frequency can be used to detect the synthesized speech – different algorithms make it unnaturally smooth, or, on the contrary – has too great of deviation. Detection of these features allow for the rejection of most artificially generated speech during verification. It should also be noted that these attacks must be worked out – impostors trying to break the system must perform many attempts by copying the parameters of the synthesized voice. The basic protocol allows the possibility of three attempts for verification. After 3 unsuccessful logins, access to the system is blocked (as is the case when using your PIN at ATMs), so the impostor is not possible to perform further attacks. An equally important obstacle for a potential impostor is that as the system response he receives only the final decision on acceptance or rejection of attempts to verification. It is not a known measure of his matches synthesized speech used to model the voice of the account holder. There is therefore no indication as to how it should modify the synthesis algorithm to increase the chances of a successful fraud. Therefore, in practice, thereby attack the system of biometrics is not a threat so important, as described in the scientific literature and the media.

Converting voice

This is a method wherein the voice of a person is converted using software in such a way that it comes close to the sound of the account holder. Modification can be:
– signal frequency characteristics,
– the course of the basic tone,
– vowel quantity – the proportions of the lengths of syllables or phonemes.

Numerous studies have shown that this method carries a certain danger, but – as in the case of speech – needs articulation, which is not feasible due to the nature of the voice biometrics system. Before the impostor can impersonate a voice effectively enough to cheat the system, the ability to log onto victim’s account will be blocked by the system.

Impostor modulation

Modulation of imitating is the simplest method of evading the speaker verification system. The impostor thus breaking into account the victim tries to imitate the account holders voice with software without additional tools. Studies show tests carried out in such a way can trigger an error effect only when impostors natural voice is extremely similar to the
voice of the victim. Of course we are talking about the similarity measured not “by ear”, but via a computer speaker recognition system to which the impostor would have access. The analysis with VoicePIN system showed the failure to access computer-based sound processing, voice imitation of 8 victims was difficult to detect by 113 respondents from real active users, was not able trick the biometric voice verification system. This means that the biometric system is based not only on the fundamental tone of voice that people perceive, but also measures the voice phenomenon over looked by people. Also, attempts to defraud the system using recordings made by professional imitators impersonating account holders were unsuccessful.

Summary
The human ear can identify a voice owner with great accuracy. But biometric voice verification systems uses advanced digital signal processing techniques that capture nuances of speech inaudible to humans.
The combination of sound analysis technology with the appropriate access management system reduces to a minimum the risk of impostor’s positive verification.