A light-based AI scans 15 deepfake videos at once, hitting 98% accuracy

A team at UCLA's Ozcan lab has built a hybrid digital-optical system that processes 15 or more video streams in a single optical pass, reporting 97.79% accuracy on a small Celeb-DF benchmark in a peer-reviewed eLight paper. The setup pairs a lightweight digital front end with a programmable spatial light modulator, a chip that shapes light to perform computation.

A standard deepfake classifier pushes every frame through digital neural-network math, one video at a time. The Ozcan group's design encodes each candidate as an optical pattern, then lets one optical pass do the discrimination work for all 15 streams at once, trading per-stream compute for light-based parallelism. The paper reports 99.86% sensitivity and 95.72% specificity on the same 15-video slice, and the lab is positioning the system as a first-pass screen that flags suspicious uploads for more expensive digital models.

The 98% figure rests on 15 Celeb-DF videos, not a field test, with no independent replication, no matched digital baseline, and no disclosed total system energy once encoding, modulation, and readout are counted. The preprint's white-box protection does not generalize to deployed security, and deepfake generators keep improving faster than any single detector. Whether the throughput advantage survives out-of-distribution video and stronger adversaries is the next question.