A video is circulating. In it, a political figure makes shocking remarks. It is shared thousands of times in an hour. Real or fake?
To find out, it has to be analysed. And when millions of videos are published every day, analysing each one costs an enormous amount of time and electricity. Researchers at the University of California, Los Angeles had a surprising idea: hand part of the work over to light.
Computing with light
An ordinary computer computes with electricity, step by step, and heats up as it does so. The researchers' system works in two stages.
First, a conventional part prepares the videos. Then the bulk of the analysis is done by sending light through specially designed layers. As it passes through them, the light is distorted in a way calculated in advance, and that simple passage performs the computation. It is physics doing the work, with almost no energy spent.
And because light can carry several images at once, the system analyses 15 videos simultaneously, in a single pass.
The results
On a collection of fake videos widely used by researchers, the system spotted deepfakes with close to 98% success. On videos it had never seen, made by Veo 3, Google's video generator, it still reached nearly 95%, after a simple adjustment.
The researchers also highlight its low consumption and its resistance to attempts to fool it. Their work was published in the scientific journal eLight.
A detector can be wrong in two ways: letting a fake through, or accusing a real one. In the published tests, the system spots almost all deepfakes, but it also wrongly classifies around 4 real videos out of 100 as fake. On the scale of a social network receiving millions of them a day, that would mean tens of thousands of real people wrongly suspected. That is why a detector must be used to sort, never to condemn.
Why light
It is no coincidence that this avenue keeps coming back. In September we reported how light could halve the energy consumption of chips. The same principle is at play here: moving information around with light rather than with electric currents costs far less. For checking billions of videos, that is a decisive argument.
What to keep in mind
This is a laboratory result, not a product. And the race between those who make fakes and those who detect them never stops: every new generator forces detectors to be retrained. For text, that race already looks lost, as we explained regarding text detectors being circumvented.
For video, this work shows at least one thing: detection can become fast and cheap enough to be done at scale. That is a necessary condition. It is not yet a solution.