UCLA: an optical processor sorts through 15 videos at a time and detects deepfakes with an accuracy of nearly 98 per cent

A lightweight digital encoder processes each video, after which light does the rest as it passes through diffractive layers: 97.79 per cent accuracy, and 94.8 per cent for Veo 3 videos.

Aydogan Ozcan's team at UCLA has developed a hybrid system, digital and optical, to detect faked videos. A lightweight digital encoder extracts spatial, spectral and temporal characteristics from each video, converts them into phase patterns displayed on a spatial light modulator; the wavefront then passes through a passive optical decoder, and pairs of detectors deliver an authenticity score for each video. Most of the calculation is therefore done by simple light propagation, which makes it possible to evaluate several videos in a single pass.

Performance

With 15 videos processed in parallel, the system achieves 97.79% accuracy, 99.86% sensitivity and 95.72% specificity; it retains 96.13% with 18 videos. Adding diffractive layers saves about 6.8 points. Compared to recent content produced by Google's Veo 3, the accuracy remains 94.8%.

The authors, who publish their work in eLight, present the tool as a first fast and energy-efficient filter: the reported videos will still have to go through heavier digital detectors for a final verdict.