UCLA's Optical Chip Detects Deepfakes at 97.79% Across 15 Streams

A UCLA group reports an optical-neural chip that spots deepfake video by shining frames through diffractive layers instead of running them through a GPU, hitting 97.79% average accuracy while screening 15 or more video streams in a single optical pass.

The paper, 'Scalable, Energy-Efficient Optical-Neural Architecture for Multiplexed Deepfake Video Detection,' appears in *eLight* and is credited to Parnian Ghapandar Kashani, Shiqi Chen and Aydogan Ozcan from UCLA's Electrical and Computer Engineering and Bioengineering departments and the California NanoSystems Institute. As summarized by ScienceDaily, the system logged 99.86% sensitivity and 95.72% specificity, held 96.13% accuracy when pushed to 18 videos in a single optical pass, and reached 94.80% accuracy on video generated by Google's VEO-3 model.

The researchers position the chip not as a replacement for digital detectors but as 'a highly sensitive first stage of a larger detection system' for content moderation and media authentication.

Because detection weights are baked into the diffractive layers themselves, the authors report 'resistance to black-box adversarial attacks' and argue the hardware-embedded parameters raise the bar against white-box ones. Adding two more diffractive layers lifted accuracy by roughly 6.8%.

Deepfake detection has been a busy lane on our deepfakes tracker, 31 pieces in the past 90 days, and almost all of it is software racing software. An optical screen, if it scales out of the UCLA cleanroom, is a different sort of answer.