Aydoğan Özcan’s UCLA Team Demonstrates Light-Based Deepfake Detection

A UCLA advance in video authentication

A UCLA research team supervised by Turkish-born professor Aydoğan Özcan has demonstrated a new way to identify manipulated videos by letting light perform part of the computation. In a laboratory experiment, the hybrid digital-optical system evaluated 15 videos simultaneously and achieved average accuracy of 97.79% on the Celeb-DF benchmark. Published in eLight on September 22, 2026, the study offers a concrete research advance in digital authentication rather than a promise that every suspicious video can now be identified. The central idea is to combine conventional artificial intelligence with a physical optical processor, distributing the work between electronic computation and light traveling through a carefully designed system.

The authors are Parnian Ghapandar Kashani, Shiqi Chen and Aydogan Ozcan, with Kashani and Chen credited as equal contributors. Their affiliations include UCLA’s Electrical and Computer Engineering Department, Bioengineering Department and California NanoSystems Institute. The paper identifies Özcan as the research supervisor and corresponding author. His connection to Türkiye is well documented: UCLA describes him as born in Istanbul, and his university profile lists an undergraduate degree from Bilkent University in Ankara, followed by graduate degrees at Stanford University. This latest project therefore brings a Turkish-American scientific connection to a distinctly American research setting, while recognizing that the experimental and analytical work was a collaborative effort.

The problem the researchers address is not simply whether an individual detector can recognize a fake. It is whether large quantities of video can be screened with a workable balance of accuracy, energy consumption and resistance to manipulation. The study explains that sophisticated digital detectors can require substantial computation, while attackers may alter videos specifically to confuse their predictions. The UCLA approach is intended as a first-stage screening system: identify potentially manipulated material efficiently, then pass suspicious cases to more demanding digital analysis. That distinction is important. The research describes a component for a larger authentication pipeline, not an independent authority capable of deciding the truth of everything shown in a video.

How light becomes part of the calculation

The process begins with digital preparation rather than an optical shortcut around all computing. A lightweight electronic encoder examines selected video frames and extracts information through spatial and frequency-domain processing. It converts the resulting representation into a phase pattern displayed on a programmable spatial light modulator. Light interacting with that pattern then propagates through the optical decoder. At the output, paired detector regions provide signals from which the system calculates a classification score for each video. In accessible terms, the electronic portion prepares the information, while the optical portion carries out part of the decision-making calculation through physical propagation. The technology is hybrid because both stages are necessary to the demonstrated result.

The ability to handle several videos together comes from dividing the optical system into multiple spatial channels. Rather than sending one complete video after another through a conventional decoding network, the researchers encode information from a batch onto the modulator and evaluate those channels during one optical pass. The paper’s main experimental configuration used 15 videos, with 12 sampled frames representing each video. This does not mean that 15 full-length videos play through a beam of light or that every frame is examined. It means that prepared representations of separate videos share the optical processing step. The researchers also arranged the detector positions and propagation geometry to reduce interference between neighboring channels.

A second configuration expanded the batch to 18 videos and used 16 frames per video. Its experimental average accuracy was 96.13%, below the 97.79% reported for the 15-video setup. The authors attribute the difference partly to increased optical cross-talk as more channels were packed into the available area. That result makes the scaling question more informative than a simple claim that more parallel processing is always better. Increasing capacity can create physical trade-offs, including less consistent performance across channels. The study suggests that a larger usable modulator aperture could support additional channels, but the demonstrated 18-video result should remain distinct from projections about much larger future systems.

The team also investigated adding passive diffractive layers to the optical decoder. These structures alter the traveling light and allow a more complex physical transformation without requiring an additional powered calculation at each layer during inference. In the researchers’ comparisons, increasing optical depth helped improve performance on more challenging material and could compensate for some reductions in the digital encoder’s complexity. This is a design option within the research, not evidence that every configuration performs equally well. The paper evaluates different combinations of encoder size, optical depth and parallel capacity, showing that the strongest choices depend on the intended balance of sensitivity, specificity and energy use.

What the test results establish

The headline accuracy figure comes from a specific blind-testing experiment using Celeb-DF, a dataset containing face-swapped and corresponding authentic videos. The researchers reserved 105 real and 105 fake videos from the official test split and kept them unseen during training. They repeated frame sampling with 100 random seeds, creating 21,000 sampled video instances from those same underlying 210 videos. That distinction prevents an inflated reading of the test size: the repeated instances assessed robustness to sampling and channel assignment, but were not 21,000 independent source videos. The trained numerical model was transferred to the physical setup without using captured experimental optical data to update its encoder weights or fine-tune its decoder.

For the 15-video configuration, the reported average sensitivity was 99.86% and specificity was 95.72%. Sensitivity describes how often manipulated videos were correctly identified; specificity describes how often authentic videos were correctly classified. The two measures matter separately because a screening system can be good at catching fakes while still flagging some genuine material. These results support the authors’ proposed role for the processor as an initial filter, with subsequent analysis handling uncertain or suspicious cases. They do not establish the probability that an arbitrary online video is fake, nor do they guarantee equivalent performance across every platform, recording condition or future generation model. The percentages belong to the stated experimental conditions.

The researchers extended their evaluation beyond face swapping by assembling a custom dataset of videos generated with Google’s VEO-3 through Gemini. After fine-tuning with 50 generated videos, the 15-channel system achieved experimental accuracy of 94.80% on the new dataset. The adaptation is notable because fully generated footage is different from inserting a face into an existing recording, and detection methods cannot automatically be assumed to transfer between those tasks. Nevertheless, the need for fine-tuning is an essential part of the result. The study demonstrates adaptation to this tested generation pipeline, not universal detection of unfamiliar synthetic media without additional examples, training or evaluation.

Energy savings, resilience and practical limits

The energy analysis contains a useful caution against treating savings in one component as savings for an entire system. With the original encoder, the authors estimate an approximately 86.5% to 95.5% reduction in decoder-level energy relative to their performance-matched digital comparison, but an approximately 12.6% to 13.9% reduction from beginning to end. The difference arises because the electronic encoder still consumes energy. In a lighter-encoder configuration, estimated total savings increased to approximately 37.8% to 41.7%, alongside changes in classification performance. These are comparisons under the paper’s stated hardware and computational assumptions, not measured electricity reductions from a commercial installation or evidence that all digital alternatives would show the same difference.

Speed also requires a more careful description than simply saying that light makes the detector faster. The paper compares an ideal digital decoder time of about one millisecond per 15-video batch with about 5.56 milliseconds for an optical decoder using a modulator refreshed at 180 hertz. A faster modulator could reduce the optical decoding time, but that comes with its own hardware considerations. The digital encoder can also limit the throughput of the combined pipeline. Consequently, the demonstrated value is parallel optical processing with favorable energy and performance trade-offs, rather than a blanket finding that the optical system always outruns electronic processing. The authors’ component-level analysis explains where future hardware changes could help.

The team tested resilience to image noise, blur, JPEG compression, physical misalignment and deliberate attempts to confuse the detector. Performance was not unchanged under every disturbance: severe degradation reduced accuracy, and different optical configurations behaved differently. The authors nevertheless report competitive robustness within their comparisons and argue that physically embedded optical parameters make some attacks harder to construct. That is a security advantage to investigate, not proof that the system is impossible to fool. Its practical role would still require monitoring, validation and an appropriate second stage. The strongest interpretation is that physical computation may provide another layer of difficulty for attackers while retaining useful detection performance under tested imperfect conditions.

For Turkish Americans following scientific contributions in the United States, the work provides a substantive example of research that connects engineering innovation with the challenge of trustworthy digital media. Its immediate achievement is an experimentally tested architecture, not an announced consumer service or a completed deployment. The paper also discloses that Özcan serves as an editor for the journal and that the three authors have a pending patent application filed through UCLA. Those disclosures belong alongside the results when assessing the research. This account explains and condenses the original open-access study by Parnian Ghapandar Kashani, Shiqi Chen and Aydogan Ozcan, published under the Creative Commons Attribution 4.0 license, while keeping its laboratory findings separate from possible future applications.