Original title: Scalable, energy-efficient optical-neural architecture for multiplexed deepfake video detection
Parnian Ghapandar Kashani, Shiqi Chen, Aydogan Ozcan · September 2026
Fake videos made by AI are spreading fast, but standard software needs massive computing power and energy to spot them. To fix this, researchers built a hybrid system that pairs a light digital front-end with an optical back-end using light waves. The system sends video data through a programmable light modulator, an optical device that shapes light to perform calculations. Because light beams can travel together without interfering, the setup scans ≥15 video streams at once in a single pass of light. It checks video authenticity accurately while cutting computational costs compared to purely digital tools. In tests on real and generated media, it correctly flagged fakes while resisting video noise, compression, setup misalignments, and adversarial attacks. By using light instead of just computer chips, the system delivers high speed, low energy use, and strong defense against digital tampering.
Imagine a teacher grading a stack of a hundred identical quizzes. Normally she reads each one line by line, which takes forever and exhausts her. But what if she could lay all the quizzes on a single sheet, shine a light through them, and a special filter marks only the wrong answers? Then she'd grade the whole stack in one go. Here, the system treats video streams like those quizzes: it places each video in its own spot on a light beam, so one pass through the beam checks all fifteen videos at once.
Today's deepfake detectors are heavy — they need large GPUs and lots of energy. This work shows that a single optical pass can handle many video streams at once, cutting computational cost and making detection scalable. That matters for platforms and regulators who need to verify authenticity in real time without blowing up energy bills.
The abstract doesn't say how the system performs on unseen deepfake types beyond the datasets tested, nor does it give a specific energy savings number. The 97.79% accuracy applies only to the Celeb-DF dataset. It also doesn't address how the hardware scales beyond 15 parallel streams or whether the digital front-end becomes a bottleneck.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Scalable, energy-efficient optical-neural architecture for multiplexed deepfake video detection — Parnian Ghapandar Kashani, Shiqi Chen, Aydogan Ozcan — eLight — doi:10.1186/s43593-026-00143-y
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.