Artificial intelligence has learned to spoof voices to fool authentication systems

Researchers at the University of Waterloo in Canada have developed a method of voice falsification using artificial intelligence to bypass voice authentication.

Here's What We Know

Voice authentication is based on the uniqueness of each person's voice. This is affected by physical characteristics, such as the size and shape of the vocal tract and larynx, as well as social factors.

Authentication systems capture nuances in voice prints. Artificial intelligence can already mimic human speech quite realistically, but the algorithms have distinctive artifacts by which analysts can identify fakes.

The technique the researchers have developed aims to eliminate these features. The idea is to "engrave" the user's voice print into the fake recording.

"Our adversarial engine attempts to remove machine artifacts that are predominant in these samples."

The researchers trained the system on speech samples from 107 speakers to better understand what makes speech sound human. To test the algorithm, they created several samples to fool authentication systems. Against some weak systems, they achieved 99% after six attempts.

However, stronger authenticators proved to be more reliable. In a test against Amazon Connect, the researchers achieved a 10% success rate in a four-second attack and 40% in less than 30 seconds.

The researchers also noted that attackers need a sample of the victim's voice as well as appropriate technical skills to carry out the attack. It's a pretty high barrier, but they urged the developers of authentication systems to work on improving the security of their technology.

Source: The Register.