The key difference is that part of the detection process takes place through the physical propagation of light. This allows many videos to be evaluated simultaneously during a single optical pass rather than requiring each one to move separately through a conventional digital processing pipeline.

The technology is detailed in the study "Scalable, Energy-Efficient Optical-Neural Architecture for Multiplexed Deepfake Video Detection," published in eLight. The researchers designed the optical AI system to serve as a high-throughput, attack-resilient first layer of defense for screening large amounts of manipulated and AI-generated video.

The Growing Challenge of Deepfake Detection

Rapid improvements in generative AI have made synthetic videos increasingly realistic, increasing the need for detection systems that are both accurate and capable of operating at large scale.

Many advanced deepfake detectors depend on enormous amounts of digital computation. A single analysis can require hundreds of billions of floating-point operations, and videos are often processed sequentially. As more content must be checked, both the processing time and energy requirements can rise proportionally.

Digital detection systems face another problem. Attackers can deliberately alter fake videos in subtle ways designed to confuse a detector and make manipulated footage appear authentic.

Professor Aydogan Ozcan and his UCLA team developed a hybrid digital-optical system intended to address both challenges.

A lightweight digital encoder first collects compact information about each video, including spatial, spectral, and temporal features. That information is transformed into a phase pattern and displayed on a programmable spatial light modulator.

The resulting optical wavefront then travels through a free-space-based, passive optical decoder. At the other end, paired optical detectors directly produce an authenticity score for each video.

In effect, the system replaces a computationally demanding digital decoding network with a physical process that can handle many streams in parallel.

Nearly 98% Accuracy Across 15 Videos at Once

In experiments using visible light, the processor examined 15 Celeb-DF videos simultaneously during each optical pass.

It achieved an average detection accuracy of 97.79%, along with a sensitivity of 99.86% and a specificity of 95.72%. Sensitivity measures how successfully the system identifies manipulated videos, making the particularly high sensitivity important for a screening tool designed to keep fake content from slipping through.

The 99.86% sensitivity translated to an average false-negative rate of ~0.14%. In other words, only a very small fraction of manipulated videos were incorrectly classified as authentic.

The researchers also pushed the system further by increasing its capacity to 18 videos in a single optical pass. Even at that level, average detection accuracy remained at 96.13%.

More Optical Layers Boost Performance

The team found that the processor could also become more capable by increasing the physical depth of its passive optical decoder without substantially increasing energy use or inference latency.

When researchers added two optimized passive diffractive layers while testing more difficult deepfake manipulations, detection accuracy improved by ~6.8%.

These phase-only diffractive layers can be manufactured as passive, static optical structures/surfaces. They perform additional calculations through the diffraction of light, meaning they do not require additional electrical power while the system is performing an inference.

That approach could allow more sophisticated processing without the same energy costs that typically accompany larger digital neural networks.

Testing the System Against Google VEO-3 Videos

The researchers did not limit their experiments to conventional face swapping deepfakes. They also challenged the processor with videos produced using Google's VEO-3 model.

Newer generative AI systems can create footage that lacks many of the obvious artifacts associated with earlier deepfake technology, making them a more difficult target for detectors trained primarily on older forms of manipulation.

With only minimal fine-tuning, the optical processor achieved 94.80% accuracy and 97.61% sensitivity on previously unseen VEO-3 videos during experiments.

The results suggest that the approach could potentially adapt as generative AI technology continues to evolve.

A Deepfake Detector Designed to Be Harder to Fool

The optical system also showed resistance to black-box adversarial attacks and provides inherent protection against white-box attacks.

Part of that security comes from the way the detector physically performs its calculations. Because some of the inference process takes place through diffraction, important parameters of the optical model are effectively embedded within the hardware.

Those parameters can be difficult for an attacker to measure, reproduce, or reverse engineer. That makes reconstructing the detector and designing carefully tailored adversarial changes that can evade it considerably more difficult.

The processor also continued working reliably when videos were affected by image noise, blur, JPEG compression, and experimental misalignments. According to the researchers, those results illustrate how optical computation could provide a more secure and dependable foundation for some artificial intelligence systems.

A First Line of Defense Against Deepfakes

Rather than replacing sophisticated digital detectors entirely, the UCLA processor is designed to work as a highly sensitive first stage of a larger detection system.

Massive volumes of video could initially pass through the parallel optical processor. Content identified as suspicious could then be sent to more computationally demanding digital models for a more detailed final assessment.

Such a system could combine the strengths of both approaches. Optical processing could provide parallel operation, low decoder energy requirements, high sensitivity, resistance to adversarial attacks, and the ability to adapt to newer AI video generators, while conventional digital systems could provide deeper analysis when necessary.

The researchers say this combination could eventually be useful for large-scale content moderation, media authentication, surveillance, and other security-critical AI applications.

The authors of this work are Parnian Ghapandar Kashani and Dr. Shiqi Chen, who contributed equally, and Professor Aydogan Ozcan. The researchers are affiliated with the UCLA Electrical and Computer Engineering Department, the UCLA Bioengineering Department, and the California NanoSystems Institute.