A hybrid optical-neural system developed at the University of California, Los Angeles can detect artificially generated deepfake videos with up to 97.8% accuracy, offering a faster and more energy-efficient alternative to conventional digital detection methods. The system, described in the journal eLight, processes 15 or more independent video streams simultaneously and is designed to serve as a high-throughput first-stage screening tool for large media platforms and data centers.
The technology arrives as generative AI models produce increasingly realistic synthetic video. Early deepfake content often contained obvious flaws such as extra fingers or garbled text, but newer diffusion models can generate highly convincing images and videos. A recent study found that humans correctly identify deepfakes only about half the time. «While AI technologies open new possibilities for content creation, they also raise growing concerns about misinformation, identity manipulation and the erosion of trust in digital media,» said Aydogan Ozcan, who led the research. «Efficient automated detection systems are therefore sorely needed.»
Existing detection systems rely on computationally intensive neural networks that are slow and consume large amounts of energy, making them difficult to scale for large-scale video processing. Deep-learning detectors are also vulnerable to adversarial attacks, in which slight alterations to fake videos can fool the system into labeling them as authentic. To address these limitations, Ozcan and his colleagues explored whether some of the computation could be performed using the physical properties of light.
In their design, a lightweight digital encoder extracts spatial and temporal optical 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 optical decoder, where paired optical detectors generate an authenticity score for each video. The researchers tested the system on face-swapped videos, real-world deepfake recordings and AI-generated videos. In one test using the Celeb-DF database, which contains identity-based face-swap manipulations alongside corresponding real videos, the system processed batches of 15 videos in a single step with an average detection accuracy of 97.79%. It achieved a sensitivity of 99.86%, meaning only 0.14% of fake videos were falsely labeled as genuine, and a specificity of 95.72%, meaning 4.18% of real videos were incorrectly judged to be fakes.
In a similar test of videos generated using Google's VEO-3 model, the optical processor reached 94.80% accuracy and 97.61% sensitivity across the 15 parallel video experiments. According to Ozcan, the result shows that after lightweight fine-tuning, the system can adapt to newer generative-AI pipelines that lack many of the artifacts found in earlier deepfakes. «Importantly, this efficiency does not come at the expense of robustness: the optical decoder is also resilient to noise, image compression and physical misalignment as well as adversarial attacks,» he said. «What makes this particularly exciting is that optical computation simultaneously provides parallelism, energy efficiency and an additional physical barrier against adversarial manipulation — properties difficult to achieve together with conventional digital hardware.»
Ozcan and his colleagues envision the technology primarily as a high-throughput, first-stage screening system for large-scale media platforms or data centers. Rather than replacing sophisticated digital forensic models, the optical processor could rapidly examine enormous numbers of videos with high sensitivity and comparatively low energy consumption, forwarding suspicious content to more computationally intensive digital models for detailed verification. Beyond deepfake detection, Ozcan said the work demonstrates that embedding AI computation within optical hardware can create new combinations of efficiency and security, since important model parameters are physically embodied in the optical system and are therefore considerably harder to extract or reproduce. Similar principles could ultimately benefit other high-throughput, security-critical AI inference tasks, he added.
The team now plans to scale up and adapt the system's architecture. Optical modulators with a larger aperture should allow substantially more videos to be multiplexed, while faster modulators such as MEMS-based phase-light modulators could further reduce latency and energy consumption per inference. Ozcan also noted that it will be important to adapt the technology rapidly so that it can handle new generations of synthetic video. The system's performance on VEO-3 generated videos, he said, is a good early indication that optical-neural systems can be efficiently updated rather than completely redesigned as generative AI evolves.
17
