An optical processor screens fifteen videos at once for deepfakes
On 1 October 2026, ScienceDaily reported that UCLA researchers led by Aydogan Ozcan had built a hybrid optical-neural processor that scores fifteen videos for manipulation in a single pass of light: a lightweight digital encoder writes each video's features as a phase pattern, and a passive optical decoder turns that light into an authenticity score. On the Celeb-DF benchmark it averaged 97.79 per cent accuracy, 99.86 per cent sensitivity and 95.72 per cent specificity at fifteen videos per pass, held 96.13 per cent accuracy when pushed to eighteen, and reached 94.80 per cent accuracy on previously unseen VEO-3 videos after minimal fine-tuning. The authors present a laboratory prototype meant as a fast first screening stage before slower digital checks, reporting that the hardware-embedded weights resisted black-box adversarial attacks and that performance held under noise, blur, compression and misalignment.
The record ends here. Everything below this line is invented.
What if the adversary attacked the room a video was filmed in, not the video?
Part one
Fifteen at Once
The flag came in at 22:41, before Amara saw the video.
That night the desk ran thirty-five hundred clips through the wire's new optical detector. Light in one end, fifteen videos at a time, a score out the other, the weights in glass nothing on the network could reach. "Unhackable remotely", the memo said.
The flag said the clip was generated. It was not. Amara had watched it four times: a night-shift worker at a chemical plant, filming a discharge pipe running a colour it should not run. Her sister Yvette worked that plant; Amara had told her all week to keep it safe.
By one, eleven clips from the same county were red, three on the desk's kill list.
The fine print gave the specificity: 95.72 per cent. Four in every hundred real videos came back marked synthetic — on a night this size, a hundred and fifty true things gone.
The explainer answered in the flat voice she distrusted: artifacts consistent with generation. The artifact sat in the same rows of pixels every time — a faint band moving up the frame. Every flagged video had been shot under the same light: the long flat panels a plant buys when it replaces sodium lamps.
At 1:40 the desk lead messaged the channel. "Kill the pipe clip. Screen says fake."
Eleven lights on the output board were red, and every one had been filmed in the same room of the same building by people who were there.
What’s real
- Fifteen videos are scored in a single optical pass.
- Celeb-DF results: 97.79% accuracy and 95.72% specificity.
- Trained weights live in the optical hardware.
- It is a prototype, pitched as a first screening stage.
What’s invented
- Amara, her sister Yvette, and the verification desk.
- Vireo V-9 panels, lot 22-C, and firmware 4.11-C.
- Ashgrove Risk and its evidentiary hardening.
- The rolling-shutter flicker that flags true video.
Part two: The Notch at 3,870
Part two isn’t on sale yet. Check back soon.
Filed under technology, deepfakes, optics, media forensics.