UCLA's optical deepfake screen classifies at least 15 videos in one pass of light

The light handles the back half of the pipeline. According to The Neuron's account of the paper, a digital encoder extracts features from sampled frames and writes them as a pattern onto a spatial light modulator ^[5]. Light then passes through an optical decoder, where propagation and diffraction perform part of the classification, and sensors turn the output intensity into one score per video ^[6]. The stream count comes from layout ^[1]. Separate regions of the optics take separate videos, so one pass classifies all of them ^[7]. Everything before the modulator runs on ordinary electronics ^[8].

The 97.79% Celeb-DF accuracy ^[2] is the exact midpoint of the 15-video run's two rates: (99.86 + 95.72) / 2 = 97.79 ^[3]. Accuracy lands on that midpoint when fake and genuine videos carry equal weight. For the figure to transfer to a platform, half of its uploads would have to be fakes. On a stream where genuine video dominates, specificity decides the result. A 95.72% specificity forwards 4.28%, about four in every hundred genuine videos, to the second stage ^[1].

The researchers pitch the processor as a first stage that flags suspicious footage before heavier digital detectors run ^[3]. I think tilting it toward sensitivity is right for that job. The stated reasoning is that a fake missed at the first stage may never be checked again, while a genuine video flagged by mistake can still go to another detector ^[11]. At 99.86% sensitivity, the filter lets through about 14 of every 10,000 manipulated videos ^[2]. Any saving then depends on how much content gets flagged and how much compute the second-stage detector needs ^[12].

The energy accounting deserves credit. The optical decoder draws roughly 1.38 to 4.11 millijoules per video in one configuration ^[15]. The September 22 paper also counts the electronics: total consumption is higher because the digital encoder accounts for much of the energy ^[4][16]. Reporting the decoder's millijoules alone would have made a better slide. Shrinking that encoder produced the energy cuts, and the smaller builds lost specificity ^[17][18]. Each point of specificity lost sends one more genuine video in a hundred to the expensive detector, so part of the first-stage saving is spent again downstream. The Neuron's write-up does not give the specificity of the lighter builds.

The Veo 3 test shows the encoder can be retrained for a new generator with a small batch, and it was scored on a held-out set of 210 videos ^[13][4]. The write-up states that the result does not show the same accuracy carrying over to an unfamiliar generator without fine-tuning ^[14]. The team also ran black-box adversarial attacks, in which an attacker can query the detector but cannot inspect its parameters ^[21].

Deployment comes down to two conditions. Total screening cost has to fall while false negatives stay within the platform's required threshold ^[19]. Hardware cost and the volume passed to second-stage detectors decide whether overall resource use drops at all ^[19].