UCLA's optical network spots deepfakes across 15 videos at once

A UCLA team has described a deepfake detector in the journal eLight whose heaviest computation is performed by light passing through passive diffractive layers. The system scores 15 video streams simultaneously at 97.79% average accuracy. The passive decoder draws no power while doing it.

Key takeaways

97.79% average accuracy across 15 videos analysed at the same time
99.86% sensitivity, meaning a 0.14% false-negative rate, and 95.72% specificity
At 18 simultaneous streams accuracy falls to 96.13%
Two extra diffractive layers lifted accuracy by roughly 6.8% with no increase in power draw
On previously unseen Google Veo 3 footage the system reached 94.80% after minimal fine-tuning

97.79%average accuracy across 15 video streams analysed at onceeLight 2026, 6(1)

Digital encoder, optical decoder

The division of labour is the point. A lightweight digital encoder pulls spatial, spectral and temporal features out of the video, then writes them as a phase pattern onto a programmable spatial light modulator. From there no silicon computation remains.

Input

Video streams

Digital stage

Lightweight digital encoder

Phase pattern on a light modulator

Optical stage

Propagation through free spaceAllow

Paired detectors read the score

The diagram shows the split of responsibilities, not the physical scale of the rig. The boundary in step three is what matters: everything past it runs unpowered, because light propagation performs the computation.

Those fixed diffractive layersDiffractive layer: A thin, static structure whose pattern bends passing light in a planned way. Once fabricated, it computes without electricity. are fabricated once and then need neither power nor reprogramming.

The parallelism comes from physics rather than chip architecture. Many streams propagate through the same decoder at once, so adding another video costs almost nothing until the system runs out of capacity.

Sensitivity over specificity

The error profile is unusual and deliberate. The two metrics measure different things, and in fake detection they do not matter equally.

\dots ▶ Symbol meaning \dots fakes correctly flagged as fakes \dots fakes let through as genuine \dots genuine clips correctly let through \dots genuine clips wrongly flagged as fabricated

A sensitivity of 99.86% means one fake in roughly seven hundred slips through, while 95.72% specificity means a share of genuine clips will be wrongly flagged as fabricated. For content moderation that is the right direction to err, but it demands a second verification stage.

Metric	Value
Average accuracy, 15 streams	97.79%
Average accuracy, 18 streams	96.13%
Sensitivity	99.86%
Specificity	95.72%
Accuracy on Google Veo 3 footage	94.80%
Gain from two extra layers	around 6.8%

The result on material from Google Veo 3, a generator the system never saw in training, is a stronger argument than headline accuracy. Deepfake detectors usually collapse when a new generator appears.

Why it matters

Deepfake detection loses today on cost rather than quality: scanning all video traffic on GPUs is economically unrealistic. Moving the decode stage into passive optics changes that arithmetic, because adding streams adds no watts. As long as the setup needs a spatial light modulator and an optical bench, though, it stays in the lab.

What's next

The paper describes a laboratory setup, with no stated plan for miniaturisation or deployment
Accuracy already slips at 18 streams, so the scaling ceiling sits close to the tested figure
Generalisation was checked against one new generator, not a broad set