This light-powered AI can spot deepfakes with nearly 98% accuracy
UCLA researchers built an AI system that uses light to analyze more than a dozen videos simultaneously, detecting deepfakes with nearly 98% accuracy. Its speed, low energy demands, and resistance to attacks could make it a powerful tool for screening the growing flood of AI-generated video.
Researchers at the University of California, Los Angeles (UCLA) have developed an optical-neural processor that uses light to accurately identify deepfake videos. Unlike traditional systems that process videos sequentially, this technology can analyze 15 or more video streams at once during a single optical pass. By incorporating the propagation of light into the detection process, the system can evaluate multiple videos simultaneously, offering a high-throughput and resilient first layer of defense for screening large amounts of manipulated video content.
The processor employs a hybrid digital-optical design, starting with a lightweight digital encoder that collects compact information about each video, including spatial, spectral, and temporal features. This information is then transformed into a phase pattern and displayed on a programmable spatial light modulator, creating an optical wavefront that travels through a passive optical decoder.
At the receiving end, optical detectors generate an authenticity score for each video, replacing a computationally intensive digital decoding network with a physical process capable of handling multiple streams in parallel. In tests using visible light, the processor identified manipulated videos from the Celeb-DF dataset with 97.79% accuracy, a 99.86% sensitivity (false-negative rate of ~0.14%), and a 95.72% specificity.
The system's performance improved slightly when tested with 18 videos per pass, reaching 96.13% accuracy. When exposed to more sophisticated deepfake manipulations, such as those generated by Google's VEO-3 model, the processor maintained an accuracy of 94.80% and a 97.61% sensitivity. The optical system demonstrated resistance to various forms of attacks, including black-box adversarial attacks and white-box attacks, due to the embedding of important parameters within the hardware and the diffraction-based calculations, which make it difficult for attackers to reverse engineer or manipulate the detector.
Written by urgent.news from ScienceDaily's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
