Counterattack launched against voice deepfakes for election meddling

It’s tough to say who is ahead in the U.S. presidential election campaign: Kamala Harris, Donald Trump, or deepfake fraudsters. With the vote still two fervid months away, the lead-up has already been haunted by deepfakes of Harris and outgoing U.S. President Joe Biden, as well as deepfake images of Taylor Swift which Trump himself circulated. It has gone far enough that Microsoft – which has teased a deepfake engine too powerful and too dangerous to release publicly – has called for deepfake fraud to be outlawed.

Call it derealizaton, enfalsification or the triumph of kayfabe: politics – and reality – are becoming more malleable by the minute. This presents both a challenge and an opportunity for digital ID providers of AI-driven tools designed to smoke out algorithmic fakery and prevent election manipulation, fraud and other nefarious acts.

Automatic real-time audio deepfake scans for videos

Reaching back to the wisdom of Sir Francis Bacon, McAfee and Lenovo are proposing a way to stop it, via the McAfee Deepfake Detector product available exclusively (for now) on Lenovo AI PCs.

“Knowledge is power, and this has never been more true than in the AI-driven world we’re living in today,” says Roma Majumder, senior vice president of product at McAfee, in an announcement on the company’s blog. “No more wondering, is this Warren Buffet investment scheme legitimate, does Taylor Swift really want to give away cookware to fans, or did a politician actually say these words? The answers are provided to you automatically and within seconds with McAfee Deepfake Detector.”

Users of Lenovo AI PCs can opt in to McAfee’s deepfake detection feature, which scans videos for AI-altered English-language audio and sends an alert within seconds if fake audio is detected. The software is trained on nearly 200,000 samples, a number that continues to grow. It leverages the neural processing units installed in Lenovo AI PCs, which allow McAfee’s AI models to perform the entire identification process on-device and off the cloud, with improved processing speed.

“McAfee does not collect or record a user’s audio in any way, and the user is always in control and can turn audio detection on or off as desired,” says the blog.

One note in the blog will surely resonate with anyone who has ever raged at choppy video or a frozen session of Baldur’s Gate. Automatic video scanning and deepfake audio detection can “significantly enhance the consumer experience,” allowing people to run deepfake detection without compromising the speed of their PC. “This ensures consumers can use their PC as usual – whether they’re gaming, browsing, or watching videos – while McAfee Deepfake Detector works quietly in the background, protecting people against deceptions and alerting them to potential scams without compromising performance.”

Pindrop forms ‘defensive grid’ in partnership with Youmail

Pindrop, which determined which TTS engine was used to create the Joe Biden audio deepfake, is also taking the AI threat to free and fair elections seriously. The voice authentication firm has partnered with YouMail to create the Election Communication Defense Grid (ECDG).

The ECDG formalizes a partnership between Pindrop and call-sensor network Youmail that has been developing for seven months. “We began collaborating with YouMail’s engineering teams the day after detecting the TTS engine behind the Biden robocall,” says a post on Pindrop’s blog.

Having now analyzed robocalls from over a thousand incumbent political candidates and their key challengers for the 2024 election, the firms have mustered under the new banner and integrated Pindrop’s Pulse Inspect deepfake detection APIs into their robocall mitigation service, which monitors billions of consumer calls across all major US carriers, to prevent fake audio robocalls from entering the conversation.

The move follows the US$100 million in debt financing that Pindrop recently secured from Hercules Capital to fund scaling of its customer base for deepfake detection and voice biometrics.

Deepfake dealers also on the financial hook for election interference

The consequences of deepfake fraud in the political arena will also extend to the conduits that distribute it. This lesson is about to cost Lingo Telecom, the voice service provider that distributed the AI-generated Biden robocalls urging voters in New Hampshire not to vote in upcoming Democratic primary elections, a million dollars. Communications Today reports that the telco has agreed to pay that sum as a fine for its role in the January robocall debacle.

The fake content was ultimately traced back to Steve Kramer, a political consultant working for a rival campaign, who admitted to hiring a street magician to create the fake audio content using a common text-to-speech (TTS) engine. Pindrop later revealed that ElevenLabs had developed the TTS in question. Kramer told CNN it took less than half an hour to create the fake voice content.

Lingo Telecom then put it out into the world – and, in doing so, unknowingly wrote a million-dollar cheque.

Kramer faces his own $6 million fine from the FCC, plus 26 criminal counts of voter intimidation and impersonating officials in New Hampshire.

Article Topics

deepfake detection  |  deepfakes  |  elections  |  fraud prevention  |  generative AI  |  McAfee  |  New Hampshire  |  Pindrop  |  United States  |  voice biometrics