UCLA's hybrid optical AI detects deepfakes 15x faster with 97.8% accuracy

The UCLA team built a hybrid deepfake video detection framework combining a lightweight digital front-end with a spatially multiplexed optical decoding back-end. By encoding multiple video streams onto a programmable spatial light modulator, the optical system performs massively parallel analog inference in a single propagation pass. This enables simultaneous processing of 15 or more videos at drastically lower energy and computational cost than purely digital deep learning models. The optical decoder uses visible light and achieves an average accuracy of 97.79%, sensitivity of 99.86%, and specificity of 95.72% on the Celeb-DF benchmark. It also proved resilient against real-world degradations like video compression, noise, experimental misalignments, and black-box adversarial attacks.

The architecture was validated across multiple datasets including classical face-swapping, real-world deepfakes, and fully AI-generated videos. The key innovation is combining optical computation with AI inference to simultaneously improve throughput, energy efficiency, and adversarial robustness—three properties that are typically trade-offs in digital systems. This approach could enable scalable, real-time deepfake detection in social media, news verification, and video surveillance without the massive GPU power required by current state-of-the-art detectors.