Deepfake Detection: How Light Tests a Video

Why in the News

Checking a deepfake video can draw not only on a computer's fast chip but also on light's physical speed and behaviour. In a study published in eLight on 22 September 2026, researchers at the University of California, Los Angeles (UCLA) presented a digital-optical system that checks 15 videos in parallel in a single optical stage. This is experimental research, not a ready public verification service. Its purpose is the initial screening of suspect videos, not issuing a certificate of final truth on every video.

The Difference Between Digital and Optical Computing

Digital computing: the processing of information on the basis of numerical instructions and bits in electronic circuits. The many mathematical operations in ordinary AI models happen this way.

Optical computing: using the spreading of light waves, their mutual interaction and measured signals for some part of a computation. In this research the digital part has not been removed; the part that extracts useful signals from the video is digital, while an important further part is done through light. So calling it a hybrid system is right.

From Video to an Optical Score

1. Choosing frames: the facial region and useful visual signals are taken from a video's selected image frames. In the research, a digital encoder combined spatial and frequency-related information. An encoder is the part that converts information into another useful form.
2. Making the phase pattern: the extracted signals are converted into a pattern representing the phase of a light wave. Phase tells which stage a wave is at in its repeating cycle.
3. Steering the light: a spatial light modulator (SLM) is a device that controls a property of an optical wave at its different locations. In this experiment the phase pattern was displayed on it.
4. Parallel decoding: the signals move ahead through light's spread in free space and through diffraction. Diffraction is the spreading of a light wave and the change in its distribution as it passes through an obstacle or structure. Pairs of detectors measure the signals for different videos, and from them a score of real or synthetic emerges.

Here multiplexing means processing the signals of several independent videos together in separate spatial channels within the same arrangement. It does not mean unlimited videos or every kind of content can be checked at once without limit. Unwanted effects between channels, the available optical area and the equipment's configuration matter.

Three Figures, Three Different Questions

In the Celeb-DF test, the average experimental results of the arrangement with 15 videos per optical stage were:

• Accuracy: 97.79%. The proportion of correct classifications across all the real and synthetic videos checked.
• Sensitivity: 99.86%. The ability to identify a synthetic video as synthetic. In this test its related average false-negative rate was about 0.14%, meaning synthetic videos slipping through the check.
• Specificity: 95.72%. The ability to identify a real video as real. This is also important for understanding the risk of false suspicion on genuine content.

Calling 99.86% the overall accuracy would be wrong. These results belong to a fixed test dataset and the experiment's conditions. In the research, repeated separate frame samplings and channel configurations were tested on a reserved set of 105 real and 105 synthetic videos. So many test runs should not be treated as that many independent new videos.

The Price of Raising Capacity

Checking 18 videos in one optical stage left the average experimental accuracy at 96.13%. This shows that rising parallel capacity and rising accuracy are not the same thing. Results also change on different, harder test content. The research showed the possibility of improving performance with additional diffractive layers in the optical decoder.

The right meaning of the energy saving: in passive optical layers, no electronic power is spent on each additional operation of the computation. Still, the digital encoder, the light source, the modulator and the detectors need energy. Calling the whole system electricity-free is not right. In the same way, real speed includes the equipment's refresh rate and the time of digital processing; merely looking at light's speed does not justify a claim of being faster than all digital systems.

Why This is Not a Final Truth Check

• The dataset's limit: a good result on one kind of face-swapping video does not guarantee the same result on every new creation technique, language and situation.
• Wrong identification: a synthetic video can slip through and a real video can be declared suspect. Both errors have social effects.
• Context is a separate question: a real video can also be shared with a wrong date, place or misleading description. A check of artificiality alone does not verify that context.
• Attacks and change: resilience was found against noise, compression and some adversarial tests, but taking it as safe from every future attack would not be proper.

Way Forward

A useful arrangement can be multi-stage: first a low-cost parallel screening, then a more detailed digital and human review of suspect content. Before real use, independent testing on varied content, open assessment of wrong identification and regular re-testing will be necessary. A technical score should be read along with the source, the original record and a check of the context. This research's big conceptual lesson is that some part of AI's computation can also be done in physical processes.

Prelims Practice Questions

Sources

• eLight: the original research, 22 September 2026
• The researchers' preprint version, 19 May 2026
• The publisher's research release, 23 September 2026

Download This Article: PDF

PCSMANTRA Editorial Team

Research and Editing

We prepare current affairs analysis, exam updates and study material in Hindi and English for UPSC, State PCS, PCS-J and other competitive examinations. Every article is verified against official sources before publication; sources are listed at the end of each article.

Read our editorial policy