Optical processor detects deepfake videos at 97.8% accuracy

By Sophie Lindqvist, Hypothesis Wire

A hybrid optical-neural system built at the University of California, Los Angeles (UCLA) detects artificially generated deepfake videos with up to 97.8% accuracy while processing 15 or more independent video streams simultaneously, and it consumes less energy than digital detectors of comparable performance. The work, led by Aydogan Ozcan, appears in the journal eLight.

The stakes for automated detection are rising. A recent study published in Nature Communications found that humans correctly identify deepfakes only around 50% of the time. Early generative models produced glaring artefacts such as extra fingers and garbled text, but modern diffusion models generate highly realistic footage.

"While AI technologies open new possibilities for content creation, they also raise growing concerns about misinformation, identity manipulation and the erosion of trust in digital media," Ozcan said. "Efficient automated detection systems are therefore sorely needed."

Why move the computation into light?

Existing detectors rely on computationally intensive neural networks that are slow and energy-hungry, which makes scaling to large-scale video processing difficult. Deep-learning detectors also remain vulnerable to adversarial attacks, in which slight alterations to a fake video fool the system into labelling it authentic.

Ozcan's team shifted part of the computation into the physical behaviour of light. The pipeline works in three stages:

• A lightweight digital encoder extracts spatial and temporal optical information from each video and converts it into a phase pattern.
• The pattern is displayed on a programmable spatial light modulator, and the encoded optical wavefront propagates through a free-space optical decoder.
• Paired optical detectors at the output generate an authenticity score for each video.

What did the benchmarks measure?

The researchers tested the system on face-swapped videos, real-world deepfake recordings and AI-generated content.

On the Celeb-DF database, which pairs identity-based face-swap manipulations with corresponding real videos, the system processed batches of 15 videos in a single step with an average detection accuracy of 97.79%. Measured figures for that test were:

• Sensitivity: 99.86%, meaning 0.14% of fake videos were falsely labelled genuine.
• Specificity: 95.72%, meaning 4.18% of real videos were incorrectly flagged as fakes.

A second experiment used videos generated with Google's VEO-3 model, which lacks many of the artefacts found in earlier deepfakes. There, the optical processor reached 94.80% accuracy and 97.61% sensitivity in the same 15-parallel-video configuration. Ozcan says this result shows the system can adapt to newer generative-AI pipelines after lightweight fine-tuning.

"Importantly, this efficiency does not come at the expense of robustness: the optical decoder is also resilient to noise, image compression and physical misalignment as well as adversarial attacks," Ozcan told Physics World. "What makes this particularly exciting is that optical computation simultaneously provides parallelism, energy efficiency and an additional physical barrier against adversarial manipulation – properties difficult to achieve together with conventional digital hardware."

The robustness claim is worth noting for procurement decisions: the specificity figure of 95.72% on Celeb-DF means roughly one in 24 real videos would be misclassified, a trade-off that matters for any pipeline triage design.

Where does the system fit in a workflow?

The team positions the technology as a high-throughput, first-stage screening layer for large media platforms and data centres rather than a replacement for sophisticated digital forensic models. In that architecture, the optical processor rapidly scans enormous volumes of video at high sensitivity and low energy cost, then forwards suspicious content to computationally intensive digital models for detailed verification.

Ozcan argues the approach has security value beyond detection. Because important model parameters are physically embodied in the optical hardware, they are considerably harder to extract or reproduce than weights stored in software. Similar principles, he says, could benefit other high-throughput, security-critical AI inference tasks.

What comes next?

The team has two concrete scaling paths. Optical modulators with a larger aperture should allow substantially more videos to be multiplexed per pass, and faster modulators, such as MEMS-based phase-light modulators, could further reduce latency and energy consumption per inference.

Ozcan adds that rapid adaptation to new generations of synthetic video remains essential, and that the VEO-3 results are an early indication that optical-neural systems can be efficiently updated as generative AI evolves rather than completely redesigned.