Quebec financial institution says AI useful to secure personal data

Desjardins - FILE PHOTO. (Daniel J. Rowe/CTV News)

Although recent reports and voices have highlighted its pitfalls and potential dangers, some financial institutions argue that artificial intelligence can also help to secure individuals' data.

Some financial institutions are using speaker recognition to authenticate customers using telephone services, based on the distinctive characteristics of their voice.

Quebec-based Desjardins introduced the technology in June 2021. Since then, 1.5 million members, who have previously consented to the process, have been identified by their voice in six million calls.

"The traditional authentication process was seen as an irritant by our members when they called," said Annie-Claude Jutras, senior manager at the Customer Relations Centre Transformation at Desjardins. "We had to spend several dozen seconds at the start of each call to complete all the steps in the process. With voice authentication, it's faster, more natural and more pleasant."

The method is also more secure, Jutras said, since Desjardins uses a passive type of authentication, where the computer system recognizes the voice during a simple conversation based on around a hundred parameters.

"We model the voice using these parameters, which give us a unique key for each member in the form of an alphanumeric code," she said. "It's like a fingerprint."

By doing this, the customer does not need to say a secret phrase to be identified, a type of authentication that has already been used by fraudsters, who recorded a person's voice without their knowledge and then personified it to gain access to personal information.

"This type of authentication takes place in an automated system," said Jutras. "Training a synthetic voice or using an artificial intelligence engine to calibrate the voice to perfection is much more difficult in the context of a fluid discussion where there is an exchange of questions and answers between the member and the Desjardins employee."

Laurent Charlin, Senior Academic Member at Mila, Associate Professor at HEC Montréal and holder of a Canada-CIFAR Chair in Artificial Intelligence, agrees.

"We humans can recognize a speaker's voice quite well, but the system can be trained to distinguish very specific intonations," he said. "If certain financial institutions have deployed this system, they must have ensured that it was better than the technology they were using before."

FAKING NOT AS EASY AS YOU THINK

According to Patrick Cardinal, professor and head of the software engineering department at the École de technologie supérieure (ETS), the technology does not yet allow us to simulate an improvised conversation using someone else's voice from a simple speech sample.

"The fact that authentication takes place during a conversation means that there is some uncertainty about the questions or exchanges to come, so it is difficult to predict the answers," he said. "What's more, to obtain something reliable, you would need a huge amount of data to train artificial intelligence: most people don't have this data online."

Cardinal is referring to voice extracts available on the internet.

The same cannot be said for television, radio or film professionals, whose voice extracts abound on the Web.

"We've already seen videos of actors or politicians saying things they've never said before, thanks to deep faking," said Charlin. "There are also applications that can use a short extract to generate your voice and make you say whatever they want."

"When in doubt, you can always fall back on traditional methods, such as asking for identification," said Jutras.

"Checking with the Canadian Anti-Fraud Centre, there have been no reports of artificial intelligence fraud using voice cloning, in particular, to access a victim's bank accounts. That said, this is something that the CAFC will continue to monitor closely," said Jeff Horncastle, Acting Customer Awareness and Communications Officer, said by e-mail.

This report by The Canadian Press was first published in French on June 11, 2023.