Researchers at the University of California, Los Angeles (UCLA) have created a novel AI system that employs light to identify deepfake videos. This optical-neural processor can analyze more than a dozen video streams concurrently, achieving an accuracy rate of 97.79% in experimental tests. The system's ability to process multiple videos in parallel during a single optical pass offers a potential solution for the increasing volume of AI-generated content.
The technology, detailed in the journal eLight, functions by offloading a significant portion of the detection process to the physical propagation of light. A digital encoder first extracts key features from video frames and converts them into a light pattern displayed on a spatial light modulator. This light then passes through an optical decoder, where its diffraction and propagation perform part of the classification. Sensors measure the resulting light intensity to generate a score for each video. This approach contrasts with conventional digital systems that typically process videos sequentially, requiring more time and energy.
The UCLA team proposes this optical system as a high-throughput, attack-resilient first stage in a larger deepfake detection pipeline. By screening large volumes of video content rapidly, it can flag suspicious footage for further, more computationally intensive analysis by digital detectors. This tiered approach aims to manage the growing challenge of deepfake proliferation, where realistic synthetic videos are becoming increasingly common and harder to distinguish from authentic content.
In experiments using the Celeb-DF benchmark, the system demonstrated an average detection accuracy of 97.79%, with a sensitivity of 99.86% and a specificity of 95.72%. High sensitivity is particularly important for a screening tool, as it minimizes the chance that manipulated videos are incorrectly classified as authentic. The researchers also tested the system's adaptability by using it on videos generated by Google's Veo 3 model. After minimal fine-tuning, the optical processor achieved 94.80% accuracy on these newer videos, suggesting potential to adapt to evolving generative AI techniques.
The system also showed resilience to certain types of attacks. The researchers tested its performance against noise, blur, and JPEG compression, finding that it remained comparatively stable. The physical nature of some optical parameters embedded in the hardware could also make the system more difficult to reverse-engineer than fully digital models.
While the optical stage consumes minimal energy, the overall system's energy use is influenced by the digital encoder. However, the researchers estimated that with a lighter encoder, the system could achieve 37.8% to 41.7% lower end-to-end energy use compared to a purely digital baseline. This reduction in energy consumption comes with a trade-off, as lighter configurations can impact accuracy and specificity.
The study remains experimental, and the researchers acknowledge that deployment would require dedicated optical hardware. The system's performance can also vary across different datasets and generative models.