We checked this claim: is it true that
an optical computing system can spot manipulated video with nearly 98 percent accuracy at high efficiency?
Laboratory tests showed 97.79% accuracy across 15 parallel video streams using the physical diffraction of light.
UCLA spots deepfakes using light waves
University engineers design an optical processor that spots fake video across parallel streams with minimal power.
In a nutshell
Engineers at UCLA have demonstrated a computing system that detects deepfake videos by bending light rather than relying entirely on electricity-hungry silicon chips. The passive optical setup processes 15 video streams at once with 97.79 percent accuracy, offering social media networks an energy-efficient way to screen massive volumes of video before handing flagged files to digital reviewers.
Highlights
- Laboratory tests recorded 97.79 percent detection accuracy across 15 simultaneous video streams.
- Accuracy remained at 96.13 percent when workload expanded to 18 parallel video feeds.
- Optical processing achieved a 99.86 percent sensitivity rate during preliminary video screening.
- The setup relies on passive light diffraction to carry out heavy mathematical computations.
Accuracy by parallel stream load
%The findingDetection accuracy dipped by 1.66 percentage points when parallel processing expanded from 15 to 18 streams.
- Benchmarked detection accuracy on the Celeb-DF dataset across 15 and 18 parallel video streams.
- Celeb-DF — A standard benchmark dataset used to evaluate deepfake detection performance
- Shows the optical architecture maintains high screening reliability under heavier concurrent streaming loads.
From the Editor’s Diary
Physical computing can break the energy bottlenecks of conventional electronic processors by turning the natural behavior of light into an instant mathematical calculation.
Who's involved
Aydogan Ozcan
Lead investigator and electrical engineering professor at UCLA
goal → Deploy passive optical networks to handle high-volume artificial intelligence screening
Parnian Ghapandar Kashani
Co-lead author and electrical engineering researcher at UCLA
goal → Prove optical hardware can catch spatial and temporal fakes in video
Shiqi Chen
Co-lead author and bioengineering researcher at UCLA
goal → Build hybrid systems that link digital encoders with optical processors
Social Media Platforms
Global video-sharing and networking networks
goal → Secure low-power screening pipelines to moderate millions of altered uploads
In short
Computer chips that crunch numbers using light instead of electricity can spot altered videos across multiple feeds at once, offering internet platforms a low-power way to filter millions of daily uploads. The breakthrough changes the math of internet moderation: instead of burning unsustainable amounts of electricity on digital computer chips, platforms can run first-round screening at the speed of light.
Major video-sharing sites will likely test optical hardware to catch computer-generated fakes before routing flagged clips to deeper digital inspection.
That prospect remains uncertain in the near term because the technology operates as a laboratory prototype rather than a commercially packaged microchip.
How it unfolded
UCLA team unveils optical deepfake detector
The UCLA team led by Professor Aydogan Ozcan formally introduced their hybrid opto-neural processing architecture in the journal eLight, accompanied by technical releases distributed via the scientific news service EurekAlert.
Technical community reviews benchmark results and stream capacity
Independent analysts and online science forums examined the published laboratory metrics, focusing on the processor's ability to screen 15 streams at once with a 99.86 percent sensitivity rate and resist digital tampering.
Where things stand
The core technical findings stand validated by peer review in the journal eLight. The hardware remains a benchtop laboratory prototype rather than a finished commercial chip. Key milestones to watch include whether engineers can shrink the free-space optical parts onto compact integrated photonic circuits, and whether detection holds against video generators newer than Celeb-DF and Google VEO-3.