A team of computer scientists at the University of Waterloo in Canada revealed that they can bypass voice authentication security systems — which let companies confirm the identity of their clients via unique voiceprints — with an almost 99% rate of success after just six tries.

For voice authentication systems, clients using remote banking, call centers and other security-critical scenarios are asked to repeat a specific phrase in their own voice as the system extracts the so-called voiceprint from the supplied phrase and stores it on a server. During later authentication attempts, the customer is required to repeat a different phrase and the features extracted from that different phrase are compared to the voiceprint saved in the system to establish the identity of the customer.

In response, malicious actors have used machine learning-enabled "deepfake" software to produce legitimate-sounding copies of victims’ voices via five minutes of recorded audio. However, developers of such voice authentication programs created so-called spoofing countermeasures to determine if the voice was human or artificially generated.

Now, the team of Waterloo researchers have created an approach that avoids these spoofing countermeasures, fooling most voice authentication systems within six attempts. To accomplish this, the researchers identified markers in deepfake audio that reveal it is computer-generated and subsequently created a program that removes those markers, thereby making it indiscernible from legitimate audio.

The team tested the program against Amazon Connect's voice authentication system, achieving a 10% success rate in one four-second attack. The rate reportedly rose to over 40% in under 30 seconds. The success rates were higher — almost 99% after six attempts — when used against less sophisticated voice authentication systems, the team reported.

"By demonstrating the insecurity of voice authentication, we hope that companies relying on voice authentication as their only authentication factor will consider deploying additional or stronger authentication measures," the researchers explained.

An article detailing the team’s work, Breaking Security-Critical Voice Authentication, was published in the Proceedings of the 44th IEEE Symposium on Security and Privacy.