LightFake Detector
UCLA's optical-neural D²NN processes 15 video streams simultaneously by physically propagating light through 5 diffractive layers — achieving 97.79% accuracywhile consuming a fraction of a GPU's energy. Published in eLight, September 2026.
Paper metrics (Celeb-DF dataset)
15 video streams in one optical pass · ~1 ns per inference
Energy per inference batch (log scale)
Optical D²NN: ~170× more efficient than GPU batch inference
How it works
UCLA researchers trained a Deep Diffractive Neural Network (D²NN) — five physical layers of finely patterned transmissive material — to perform feature extraction on video frames using only the natural propagation of coherent light. No inference-time power is consumed in the diffractive layers themselves: the physics handles it for free.
Spatial multiplexing tiles 15 video channel inputs side-by-side across the input plane. A single optical pass through the layers simultaneously produces 15 detector responses at the output plane. This is analogous to running 15 neural network inferences in perfect parallel with zero additional energy cost.
The system exploits the fact that deepfakes introduce subtle spatial-frequency artifacts from GAN/diffusion upsampling — artifacts that the trained diffractive layers learn to scatter toward the "fake" detector pixel while letting authentic frames converge on the "real" pixel.
Detection pipeline
Based on Scalable, Energy-Efficient Optical-Neural Architecture for Multiplexed Deepfake Video Detection, eLight (Springer), 2026. Celeb-DF v2 benchmark. Energy figures are illustrative order-of-magnitude comparisons.