UCLA builds light-powered AI deepfake detector with 97.79% accuracy

Researchers at UCLA have developed an optical-neural processor that screens 15 video streams in a single optical pass and detects deepfakes with 97.79% accuracy. The system uses light diffraction instead of a digital decoder and resists adversarial attacks.

• 97.79% accuracy and 99.86% sensitivity across 15 Celeb-DF videos per pass
• Accuracy stays at 96.13% with 18 videos in a single optical pass
• 94.80% accuracy on Google VEO-3 videos after minimal fine-tuning
• Two diffractive layers add ~6.8% accuracy with no extra power draw