The Detector That Uses Light to Catch a Fake

By Ellis Ward, Resident Expert, AI Systems & Research Norms

A research team at UCLA has built a deepfake detector that processes video using light rather than conventional computing circuits. The system, described in a paper published in the Springer journal eLight, can screen more than fifteen video streams in a single optical pass and correctly identified manipulated content around 97.8% of the time on a standard benchmark dataset. Those numbers are worth paying attention to, but so are the conditions under which they were produced.

What the system actually does

The detector is a hybrid architecture. One part is digital: a lightweight encoder that compresses the visual information from a video frame into a compact representation. The second part is optical: a passive decoder built around a spatial light modulator, a component that shapes and routes light rather than running calculations through transistors.

The key word is passive. The optical decoder does not consume the power that conventional compute hardware does because it is not executing operations in the usual sense. Light passes through the modulator, which has been configured to sort and separate the encoded signals, and the resulting pattern carries the classification output.

Because the optical path can handle many signals simultaneously, the team led by Prof. Aydogan Ozcan was able to run fifteen or more video streams through a single optical pass. That parallelism is the architectural advantage the paper is built around. Conventional deepfake detectors run sequentially, one video at a time, and typically require hundreds of gigaFLOPs per inference. The UCLA system moves that computational burden off the digital processor and onto photons.

The paper is authored by Kashani, Chen, and Ozcan, and published in eLight under Springer.

The numbers, and what they mean

On the Celeb-DF benchmark, a widely used deepfake detection dataset, the system reported 97.79% accuracy, 99.86% sensitivity, and 95.72% specificity when processing fifteen simultaneous video streams. When scaled up to eighteen streams, accuracy remained at 96.13%.

It helps to know what those three metrics measure separately.

Sensitivity measures how often the system correctly flags a fake. At 99.86%, it is very rarely missing manipulated content. That is the property that makes it useful as a first-stage screener: you do not want fakes slipping through undetected.

Specificity measures how often the system correctly clears real content. At 95.72%, it is flagging roughly 4 in 100 authentic videos as suspicious. Those are false positives. In a screening context, false positives create work downstream but do not spread harmful content. Whether that tradeoff is acceptable depends on the volume being processed and what resources exist for secondary review.

Accuracy is the overall correct classification rate across both categories. The 97.79% figure reflects the weighted combination of those two.

The team also tested the system against video generated by VEO-3, a recent AI video generation model. On that content, accuracy was reported at 94.80%. The lower figure compared to Celeb-DF is expected: VEO-3 video differs in how it was produced, and a system trained primarily on one distribution will generally be less precise when the input shifts. The 94.80% figure is relevant because it reflects something closer to current conditions.

For comparison, the paper cites research showing that humans identify AI-generated images at roughly 52 to 55 percent accuracy, just above chance. That baseline matters for understanding what a functional automated screener is being measured against.

How it holds up under stress

A common weakness in detection systems is brittleness. A detector that performs well on clean, well-lit, stable video may fail on content that has been compressed, blurred, or subjected to minor geometric shifts. The paper reports that the optical-neural system remained robust under several conditions: Gaussian noise, blur, JPEG compression artifacts, and moderate spatial misalignment.

The team also notes that conventional digital detectors are vulnerable to adversarial attacks, where small, deliberate perturbations to an image or video are designed to fool the classifier. Optical processing has a different error profile. Because the manipulation happens at the physical layer, certain adversarial strategies that work against digital classifiers do not transfer cleanly. This is an interesting property, though it has limits. The paper does not claim adversarial immunity, and a determined actor working specifically against an optical system would approach the problem differently.

What this is not

The paper is a lab result. The authors are explicit that this is a high-sensitivity first-stage screener, not a replacement for deeper analysis. A video flagged by the optical system still needs a secondary digital review before any consequential decision is made. The value of the system is in the throughput: it can process many streams quickly and cheaply, which is where a first filter is most useful. The detailed verification happens after.

There is no deployment described here. The system has not been integrated into a platform, a content moderation pipeline, or any production environment. The Celeb-DF dataset, while standard, represents a specific distribution of manipulated content. Real-world deepfakes vary in method, quality, and origin, and performance on a benchmark does not automatically translate to performance across that range.

The move from 97.79% at fifteen streams to 96.13% at eighteen streams is small but visible. It suggests that accuracy is not constant as parallelism increases, which is a parameter worth understanding before any real deployment.

One further note on dataset scope: Celeb-DF contains celebrity video manipulations. Content that looks different in texture, lighting, or compression from that dataset may not behave the same way under classification.

Why the approach matters

Deepfake detection has largely followed the same pattern as deepfake generation: digital models competing in an escalating cycle. A detector is trained, generators adapt, the detector is updated, and so on. Adding an optical processing layer changes the nature of that competition in a meaningful way. Adversarial strategies developed for digital classifiers do not transfer directly to a system where the classification is partly a function of how light physically propagates.

That does not solve the detection problem. But it opens a different design space, particularly for scenarios where throughput is the constraint. A broadcaster screening live video feeds, a platform reviewing uploads at scale, or an infrastructure operator monitoring many streams simultaneously all face versions of that problem. Processing fifteen-plus streams in a single optical pass is a substantively different capability than running the same streams through sequential digital inference, even if the downstream verification step remains digital.

Energy efficiency follows from the same architecture. Optical components do not consume power the way compute hardware does during inference. In a context where AI infrastructure energy costs are increasingly a real planning concern, that property has practical relevance beyond the lab.

The practical takeaway

The UCLA system demonstrates that optical processing is a viable path for real-time, parallel deepfake screening at meaningful accuracy levels. The 97.79% figure on Celeb-DF and the 94.80% figure on VEO-3 content are both strong, and the reported robustness to noise and compression suggests the underlying method is not fragile.

It is not a finished solution to deepfake detection. The benchmark is controlled. Deployment is not demonstrated. The sensitivity-specificity tradeoff, while favorable for a screener, produces false positives that require downstream handling. And the research captures one moment in a rapidly changing landscape: both generation and detection capabilities continue to develop.

The paper contributes a concrete demonstration that a different computational substrate, one that uses light at the classification stage, can perform at competitive accuracy while handling multiple streams simultaneously. That is a meaningful result, and one that will likely influence how detection infrastructure gets designed as scale becomes the dominant engineering constraint.

Sources and further reading

Kashani, Chen & Ozcan. Scalable, energy-efficient optical-neural architecture for multiplexed deepfake video detection. eLight, Springer. https://link.springer.com/article/10.1186/s43593-026-00143-y

This light-powered AI can spot deepfakes with nearly 98% accuracy. Rocket News, October 2026. https://rocketnews.com/2026/10/this-light-powered-ai-can-spot-deepfakes-with-nearly-98-accuracy/