Engineers at UCLA led by Professor Aydogan Ozcan have built a deepfake screener that does part of its thinking with light instead of software. On Celeb-DF face-swap videos, tested 15 at a time in a single optical pass, the system reached 97.79% average accuracy and 99.86% sensitivity. Against Google’s Veo-3, a generator it had not been trained on, accuracy came in lower, at 94.80%.
Those are two separate results from two separate tests, and the distance between them is the honest centre of the work. A single accuracy figure on a well-known benchmark describes how the device behaves on that benchmark, which is a narrower claim than it first sounds.
A spatial light modulator as the classifier
The design is a hybrid. A lightweight digital encoder pulls compact spatial, spectral and temporal features out of each video, and those features are written as a phase pattern onto a programmable spatial light modulator. The modulated light wave then travels through a passive, free-space decoder made of phase-only diffractive layers, and a pair of optical detectors reads out an authenticity score for each clip, according to the UCLA release carried by ScienceDaily.
Because the decoding happens as light propagates through the stack of layers rather than being computed step by step on a chip, many videos can share the same hardware at once, and the diffractive layers draw no electrical power during inference. That is the selling point: throughput and energy, not a new kind of forensic clue. The paper, published in eLight by Parnian Ghapandar Kashani, Shiqi Chen and Ozcan, describes the optical stage as a screening layer that sits behind a small digital front end, and the Optica news summary says the team presents it as a fast first pass for online content providers and credits it with roughly one-tenth the energy of the best digital detector it was compared against.
What 97.79% on Celeb-DF actually measures
The preprint abstract pins the 97.79% figure to a specific setup: Celeb-DF, 15 videos in parallel, averages of 97.79% accuracy, 99.86% sensitivity and 95.72% specificity. Sensitivity is the share of fakes caught, and it is the number that reads best. Specificity is the share of genuine videos correctly cleared, and at 95.72% it is noticeably lower, which means real footage is the thing more likely to be wrongly flagged. The figure is overall accuracy, not recall on fakes and not an AUC.
The release does not state the total number of videos in the test set, and Celeb-DF is the only dataset it names. The preprint says other datasets covering classical face swaps, real-world deepfakes and fully AI-generated video were also used, without naming them in the abstract. Packing more clips into one pass costs accuracy: at 18 videos per pass the average fell to 96.13%, so that accuracy depends on the 15-video configuration.
The Veo-3 result is the more interesting one for anyone worried about modern fakes. Those clips were previously unseen, and with minimal fine-tuning the system scored 94.80% accuracy and 97.61% sensitivity. Google’s Veo-3 produces fully synthetic footage rather than a swapped face, so a drop from 97.79% to 94.80% is a modest price for crossing from one kind of fake to another, though the two tests used different samples and are not strictly comparable. Veo-3 is the only modern text-to-video generator the release names, so performance on other current generators remains unreported, and the Veo-3 sample size is not given either.
The arms race a benchmark cannot settle
The authors argue that a physical decoder is harder to attack than a purely digital one. They conclude, in the words of the paper, in favour of “hybrid optical-electronic computing as a scalable and physically secure paradigm for trustworthy AI systems,” and the preprint says the system was tested against degradation, noise, compression, misalignment and black-box adversarial attacks. The release describes that robustness only qualitatively, with no attack methods or numbers, so the security claim cannot yet be weighed against an adversary who adapts.
Detectors and generators improve together, which is why an adaptive adversary is the real test. The physical design may make the hardware itself harder to probe, yet the encoder in front of it is still software, and software can be studied. A forger who can query a screener, or who simply trains against public detectors, can learn what it flags. Earlier ScienceDaily coverage of Google’s own deepfake-hunting tool and of work on spotting deepfakes through the eyes shows the same pattern: each method catches the fakes of its day, and each is then measured against newer ones.
The sensible reading is the one the authors themselves offer, and it matches how the release frames the project as a first-stage screener rather than a standalone verdict machine. The optical stage is a cheap first filter that triages a large volume of video, and anything it flags still needs heavier digital analysis. A 97.79% score on face swaps and a 94.80% score on one unseen generator show the idea works under controlled conditions. They do not show how it would fare on compressed social-media uploads, on generators released next year, or against someone trying to beat it.
This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.
More from Morning Overview
