Researchers have demonstrated a deepfake detector that uses light to help assess multiple videos simultaneously, offering a possible route to less costly screening of synthetic media.
Published in eLight on September 22, the study combines a digital neural network with an optical processor. Software extracts information from video frames and converts it into patterns displayed on a programmable light modulator. Light propagation then performs part of the classification work.
In laboratory experiments using the Celeb-DF dataset, the system processed 15 videos in each optical pass and achieved average accuracy of 97.79%. Its sensitivity to fake videos reached 99.86%, while specificity for genuine videos was 95.72%.
The experimental test used 210 held-out videos, repeatedly sampled with different frames and batch arrangements. Those repetitions tested robustness but did not represent thousands of distinct videos.
The researchers also examined other datasets and compared optical decoding with digital alternatives. Their energy analysis indicates potential savings at the decoder stage; digital preprocessing still consumes energy, an important distinction in low-energy optical computing.
The proposed role is initial screening, with suspicious material passed to more sophisticated systems. Wider deployment would require validation beyond these controlled tests.