A detector that uses light instead of electronics for part of its work can check 15 videos for deepfakes at the same time. Engineers at UCLA, led by Aydogan Ozcan, describe it in the journal eLight, published on October 1.
Checking every uploaded video with a large digital model takes a lot of computing power and energy. The UCLA design moves the heaviest step into optics. Light does the calculation as it travels, so many videos can be processed in one pass.
Half digital, half light
A small digital encoder first pulls features from each video: what is in the frame, its colors and how it changes over time. It turns them into a pattern on a device that shapes light. The light then passes through passive layers that work as the decoder, and pairs of light detectors give each video a score for real or fake.
On a standard set of celebrity deepfakes, the system checked 15 videos per pass with 97.79 percent average accuracy, the paper reports. With 18 videos at once it reached 96.13 percent. On videos made with Google's Veo 3, which it had never seen, it scored 94.8 percent.
A first filter
The authors do not present it as a full replacement for digital detectors. They see it as a fast, sensitive first stage that sends suspicious videos on to bigger models.
The tests ran in a lab, on known datasets. The idea is that cheap light does the sorting, and expensive computing only looks at what is left.
