Researchers at the University of California, Los Angeles (UCLA) have developed a scalable optical-neural processor that can detect deepfake videos with high accuracy while analyzing 15 or more video streams simultaneously. Instead of relying entirely on conventional digital hardware to process videos sequentially, the system offloads a major stage of the detection pipeline onto physical light propagation, enabling multiple videos to be screened in parallel in a single optical pass.
As described in the study “ Scalable, Energy-Efficient Optical-Neural Architecture for Multiplexed Deepfake Video Detection ,” published in eLight , the optical AI framework is designed as a high-throughput, attack-resilient first line of defense for identifying manipulated and AI-generated videos at scale.
The rapid proliferation of realistic AI-generated content has created an urgent need for reliable and robust deepfake detection. Many state-of-the-art detectors require hundreds of billions of floating-point operations per inference and analyze videos sequentially, causing their latency and energy consumption to grow proportionally with the volume of content being screened. Conventional digital detectors can also be vulnerable to adversarial attacks specifically engineered to make fake videos appear authentic to the detection system.
The UCLA team, led by Professor Aydogan Ozcan, addressed these challenges through a hybrid digital-optical architecture. A lightweight digital encoder extracts compact spatial, spectral, and temporal information from each video and converts it into a phase pattern displayed on a programmable spatial light modulator. The encoded optical wavefront then propagates through a free-space-based, passive optical decoder. At the output, paired optical detectors directly generate an authenticity score for each video, replacing a computationally demanding digital decoding network with a parallel physical process.
In a visible-wavelength experimental demonstration, the optical system analyzed 15 Celeb-DF videos simultaneously during each optical pass and achieved an average detection accuracy of 97.79%, sensitivity of 99.86%, and specificity of 95.72%. Its near-perfect sensitivity corresponds to an average false-negative rate of ~0.14%, an important advantage for a first-stage screening system intended to minimize the number of manipulated videos that escape detection. Even when the multiplexing capacity was increased to 18 videos per optical pass, the experimental system maintained an average accuracy of 96.13%.
The researchers further demonstrated that increasing the physical depth of the passive optical decoder can improve its performance without substantially increasing its energy consumption or inference latency. On more challenging deepfake video manipulations, adding two optimized passive diffractive layers increased detection accuracy by ~6.8%. Because these phase-only diffractive layers can be fabricated as passive, static optical structures/surfaces, they perform additional computations through diffraction without requiring additional electrical power during inference.
To challenge the system beyond conventional face-swapping benchmarks, the researchers also tested it on videos generated using Google’s VEO-3 model. With only minimal fine-tuning, the optical processor experimentally achieved 94.80% accuracy and 97.61% sensitivity on previously unseen VEO-3 videos. This result demonstrates the framework’s potential to adapt to newer generative-AI pipelines that lack many of the artifacts found in earlier deepfakes.
The optical processor also demonstrated resilience to black-box adversarial attacks and offers inherent protection against white-box attacks. Because part of the inference is physically performed via diffraction, critical parameters of the optical model are embedded in the hardware and are difficult to measure, replicate, or reverse-engineer. This physical protection makes it substantially more challenging for an attacker to reconstruct the detector or develop adversarial perturbations that evade it. The processor also remained robust to image noise, blur, JPEG compression, and experimental misalignments, highlighting the potential of optical computation as a secure and trustworthy foundation for artificial intelligence.
The optical processor is intended to operate as a highly sensitive first-stage screening system. Large volumes of videos could first be examined efficiently by the parallel optical processor, with flagged video content subsequently forwarded to more computationally intensive digital models for final verification. This combination of parallelism, low decoder energy, high sensitivity, adaptability to new generative models, and strong adversarial resilience could support large-scale content moderation, media authentication, surveillance, and other security-critical AI applications.
The authors of this work are Parnian Ghapandar Kashani and Dr. Shiqi Chen, who contributed equally, and Professor Aydogan Ozcan. The researchers are affiliated with the UCLA Electrical and Computer Engineering Department, the UCLA Bioengineering Department, and the California NanoSystems Institute.
eLight
energy-efficient optical-neural architecture for multiplexed deepfake video detection