As AI video becomes easier to produce, platforms face higher costs for verifying large volumes of uploads. Running footage through computationally intensive detectors requires processing time and energy. To reduce some of that demand, UCLA researchers are testing whether optical hardware can handle part of the classification process.
Their hybrid optical-neural processor can examine information from at least 15 video streams during a single optical pass. In experiments on the Celeb-DF benchmark, the system reached 97.79% accuracy. The researchers propose using it as an initial screening stage that flags suspicious footage before more computationally intensive digital systems perform further verification.
The study, published September 22 in eLight, remains experimental. Its main question is whether optical computation could reduce the processing required for large-scale deepfake screening.
How light classifies a video
The process starts with conventional computing. A digital encoder extracts features from sampled video frames, then converts those features into a pattern displayed on a programmable device called a spatial light modulator.
Light passes through an optical decoder, where propagation and diffraction perform part of the classification computation. Sensors measure the resulting light intensity and use those measurements to produce a score for each video. Since separate regions of the optical system process different video inputs, multiple videos can be classified during the same optical pass.
The optical stage therefore handles information already prepared by a digital encoder, so the complete system still relies on electronic computation.
Why sensitivity matters in first-stage screening
In the 15-video Celeb-DF experiment, the system achieved 99.86% sensitivity, which measures the share of manipulated videos correctly flagged. Its 95.72% specificity measures the share of genuine videos correctly recognized. Both results apply to this particular benchmark and system configuration.
High sensitivity matters during the initial screening stage because a manipulated video that slips through may never be checked again, while a genuine video flagged by mistake can still be reviewed by another detector. The researchers therefore position the optical system as a filter that prioritizes catching suspicious content before a more computationally intensive system evaluates it.
For platforms, this could reduce the number of uploads sent through full forensic analysis. Actual savings would depend on the share of flagged content and the computing resources the second-stage detector requires.
The researchers also tested videos generated with Google’s Veo 3. After fine-tuning the digital encoder with 50 generated videos, the system reached 94.80% experimental accuracy on a held-out set containing 105 real and 105 generated videos. The result shows that the system can adapt to Veo 3 after limited additional training, but it does not show that the same accuracy would carry over to an unfamiliar generator without fine-tuning.
The energy calculation includes the digital work
The paper estimates optical-decoder energy consumption at approximately 1.38–4.11 millijoules per video in one configuration. Total system consumption is higher because the digital encoder accounts for much of the system’s energy use.
With a lighter encoder, the authors estimated 37.8%–41.7% lower end-to-end energy use than the study's digital baseline. The lighter configurations, however, also reduced accuracy and specificity. The lower energy use therefore came with a measurable performance trade-off and should be considered separately from the headline 97.79% accuracy result.
Real-world deployment would depend on whether the system lowers total screening costs while keeping false negatives within the platform’s required threshold. Hardware costs and the volume of content passed to second-stage detectors would also determine whether the approach reduces overall resource use.
The resilience tests have limits
The researchers tested the system against noise, blur, and JPEG compression. They also evaluated physical misalignment in the optical setup. Detection performance remained comparatively stable under several tested conditions, although more severe degradation reduced accuracy.
The system was also evaluated against black-box adversarial attacks, where an attacker can query the detector but cannot inspect its internal parameters. The authors argue that some optical parameters are embedded in the physical hardware, which could make the system harder to reconstruct than a fully digital model.
The results show resistance to the black-box attacks included in the study, but they do not establish how the detector would perform against other attack methods or future techniques.
Detection and provenance answer different questions
Provenance systems such as C2PA Content Credentials and Google’s SynthID approach authenticity from another direction. When valid credentials are present, Content Credentials provide cryptographically signed, tamper-evident information about a file’s origin and editing history. SynthID embeds invisible watermarks in AI-generated media, including video, and provides tools that can check for those signals.
A content-based detector instead examines the footage itself for evidence of synthesis or manipulation. It can serve as another verification method when provenance data is unavailable or has been removed. The UCLA study does not describe its detector as a SynthID verification tool.
For now, the system remains a laboratory prototype rather than a commercially deployed detection service. Its performance varies across datasets, and deployment would require dedicated optical hardware. The authors also disclose a pending patent application filed through UCLA.
The next test is whether optical screening can maintain its accuracy and energy advantages under platform traffic and hardware constraints. Further testing would also need to address attack methods outside the laboratory evaluation.
