Deepfakes are getting better, faster, and cheaper to make. Every day, platforms are flooded with videos that may or may not be real, and checking them all with traditional software is expensive and slow. A research team at the University of California, Los Angeles (UCLA) believes part of the answer may come from an unexpected place: light itself.
The team has developed a hybrid system that uses the physical behavior of light to help analyze video, detecting deepfakes with close to 98% accuracy while screening more than a dozen videos at the same time. Here is how it works and why it could matter for the future of online trust.
The Growing Deepfake Problem
Generative AI has made synthetic video remarkably convincing. Faces can be swapped, voices cloned, and entire scenes created from a text prompt. Newer video models produce footage that is increasingly difficult for people to identify as fake.
Detection tools exist, but most run entirely on digital hardware such as GPUs. They typically process videos one after another, which becomes costly when a platform needs to screen enormous volumes of uploads. Many detectors can also be fooled by small, deliberate changes to a video.
The challenge is not only accuracy. It is accuracy at scale, at a cost platforms can actually afford.
How the Light-Powered Detector Works
The UCLA system, led by Professor Aydogan Ozcan, combines conventional digital processing with an optical component. The process works in several stages:
- Digital encoding: A digital encoder extracts important features from sampled frames of each video.
- Combining videos: Features from multiple videos are merged into a single pattern.
- Light modulation: That pattern is displayed on a programmable device called a spatial light modulator, which shapes a beam of laser light.
- Optical decoding: The light travels through passive diffractive layers, and the way it spreads and bends performs part of the classification work.
- Scoring: Sensors measure the resulting light intensity and produce an authenticity score for each video.
The key idea is that some of the computing happens simply through light moving through space. That step requires very little energy and can handle many signals in parallel.
The Results: Strong Accuracy at Scale
The research was published in the journal eLight under the title “Scalable, Energy-Efficient Optical-Neural Architecture for Multiplexed Deepfake Video Detection.” Testing used videos from Celeb-DF, a well-known deepfake benchmark dataset.
Key findings include:
- 97.79% average accuracy while screening 15 videos at once.
- 99.86% sensitivity, meaning it caught almost all of the fake videos.
- 95.72% specificity, meaning it correctly cleared most genuine videos.
- 96.13% accuracy when the system was pushed to 18 videos in a single optical pass.
When researchers tested more difficult manipulations, adding two optimized diffractive layers improved accuracy by roughly 6.8 percentage points, from about 89% to nearly 96%. According to Optics & Photonics News, that enhanced setup outperformed the best digital detector tested while using around ten times less energy.
Tested Against Modern AI Video
A detector that only works on older deepfakes would have limited value. The team therefore also tested the system on videos created by Google’s VEO-3 video generation model, which produces highly realistic synthetic footage.
With only minimal fine-tuning, the system reached 94.80% accuracy and 97.61% sensitivity on VEO-3 videos it had not seen before. The processor also continued to function under common real-world problems such as image noise, blur, JPEG compression, and small misalignments in the optical setup.
Why Light Has an Advantage
Optical computing offers three main benefits over purely digital systems:
- Parallelism: Many video streams can be combined into one beam of light and analyzed together.
- Speed: Light-based processing happens extremely quickly as the beam passes through the optical layers.
- Energy efficiency: Passive optical layers do not need electricity to perform their part of the computation.
The researchers see the system as a high-throughput “first line of defense.” It could quickly screen huge volumes of video and flag suspicious content, passing only those clips to slower, more detailed digital analysis.
Limitations to Keep in Mind
As promising as the results are, the system is still a laboratory prototype rather than a commercial product. Lighter configurations that used even less hardware also reduced accuracy, and more severe image degradation lowered performance.
The next big test is whether optical screening can maintain its accuracy and energy savings under the messy conditions of real platform traffic and hardware constraints.
Detection is also only one part of the solution. Other approaches, such as Google’s SynthID, embed invisible watermarks into AI-generated media so its origin can be checked later. Detection asks whether a video looks fake; provenance asks where it came from. Both are likely to be needed.
Frequently Asked Questions
Can I use this deepfake detector today?
No. It is a research prototype developed at UCLA and has not been released as a commercial service.
How accurate is it?
In testing on the Celeb-DF dataset, it achieved 97.79% average accuracy while analyzing 15 videos at once.
Does it work on the newest AI videos?
The team tested it on videos generated by Google’s VEO-3 model and reported 94.80% accuracy after minimal fine-tuning.
The Bottom Line
UCLA’s light-powered detector shows that the fight against deepfakes may not be won by software alone. By letting physics handle part of the work, researchers achieved strong accuracy while screening many videos at once and using far less energy. If the approach can scale beyond the lab, it could become a valuable tool for platforms trying to keep up with an ever-growing wave of synthetic media.