Deepfake makers can now evade an unusual detection method

Deepfake videos that feature digital manipulations of people’s facial expressions and voices can also depict realistic heartbeats, making them even harder to spot.

“We now know that just because a person in a video has a measurable pulse, it doesn’t mean that we can assume they are real,” says Hany Farid at the University of California, Berkeley, who was not involved in the research.

This development comes as deepfakes that have been digitally altered or generated by artificial intelligence are ensnaring celebrities and ordinary people alike in convincing but false pornography, financial scams and political propaganda. Previously, researchers had experimented with spotting deepfakes by identifying changes in skin colour related to blood flow and heart rate, but this research shows that some deepfake videos can still present a passable pulse.

Peter Eisert at the Fraunhofer Institute for Telecommunications in Germany and his colleagues developed a deepfake detector that could analyse the pulses of people in genuine and deepfake videos. They also filmed a new set of genuine videos featuring a dozen people’s facial expressions, while simultaneously recording participants’ heart rates so that they could verify the accuracy of their detector.

Then, the researchers inserted digitally altered faces into their genuine videos – a move that should have alerted their deepfake detector. Instead, they found that the detector perceived realistic pulses in both the fakes and the original videos.

“Just because one or a few deepfake generators can reproduce this physiological signal, it doesn’t mean that all deepfake generators can,” says Farid.

The team has already begun experimenting with new ways to spot deepfakes, such as identifying local blood flow patterns in people’s faces. But such methods may have a “limited shelf life”, says Siwei Lyu of the University at Buffalo in New York who was not involved in the work. That’s because new generative AI tools can more convincingly mimic realistic heartbeats and other physiological signals, and it can be difficult to extract a heart rate signal from low-quality videos.

Instead, the most effective detection techniques attempt to identify more subtle differences between genuine and deepfake videos, such as image pixel brightness, that are “non-intuitive to human viewers”, says Lyu.