UCLA researchers have developed a new deepfake detection system that uses light-based optical computing to identify manipulated videos. In experiments, the system achieved 97.79% average detection accuracy while processing 15 videos simultaneously, offering a potential alternative to conventional detectors that rely heavily on digital computation.
The research, published in the journal eLight on September 22, 2026, combines a digital AI encoder with an optical processor. Rather than relying only on digital analysis of video frames, the system converts extracted information into light patterns and uses physical light propagation to help determine whether footage is authentic or manipulated.
Key Takeaways
- UCLA researchers have developed a hybrid optical-neural system for detecting deepfake videos.
- The technology achieved 97.79% average accuracy when processing 15 videos simultaneously.
- Its sensitivity reached 99.86%, while specificity was 95.72% in the reported experiment.
- The system can process up to 18 video streams in a single optical pass, with 96.13% average accuracy.
- It uses light diffraction and optical detectors to perform part of the AI computation.
- Researchers also tested the system on videos generated using Google’s Veo 3.
- The technology is designed as an initial screening system, not a complete replacement for conventional deepfake detection tools.
How Does the UCLA Deepfake Detection System Work?
The technology was developed by researchers at the University of California, Los Angeles, under Professor Aydogan Ozcan.
The research team includes Parnian Ghapandar Kashani and Shiqi Chen, who contributed equally to the project.
The system combines conventional digital processing with optical computing. This approach allows it to extract information from videos digitally before using physical light propagation to perform part of the detection process.
The process has three main stages.
1. Digital feature extraction
A lightweight digital encoder first examines selected video frames and extracts important information.
This includes spatial, spectral and temporal features. These characteristics help the system understand patterns in the video without requiring it to process every visual detail through a large digital neural network.
2. Converting information into light patterns
The extracted information is converted into a phase pattern and displayed using a programmable spatial light modulator.
This device controls the properties of incoming light to create an encoded optical wavefront.
Multiple video streams can be combined into the optical processing system, allowing them to be evaluated in parallel.
3. Optical processing and authenticity scoring
The encoded light travels through a passive optical decoder using free-space propagation and diffraction.
At the output, paired optical detectors measure differences in light intensity and generate an authenticity score for each video.
The system then uses these scores to identify potentially manipulated footage.
The key difference is that part of the computation takes place through the physical behaviour of light rather than being performed entirely by conventional digital processors.
System Achieves 97.79% Accuracy Across 15 Videos
The researchers tested their technology using the Celeb-DF dataset, which contains real videos and manipulated videos created through face-swapping techniques.
In the experiment, the optical processor examined 15 video streams simultaneously.
It achieved the following results:
| Performance metric | Reported result |
|---|---|
| Average detection accuracy | 97.79% |
| Sensitivity | 99.86% |
| Specificity | 95.72% |
| Simultaneous video streams | 15 |
| Accuracy with 18 video streams | 96.13% |
Sensitivity measures how effectively the system identifies manipulated videos. Its 99.86% sensitivity means that approximately 0.14% of manipulated samples were incorrectly classified as authentic in the reported test.
Specificity measures how effectively the system recognises genuine videos. The reported 95.72% specificity indicates that some authentic videos were incorrectly flagged as manipulated.
This distinction matters because a deepfake detector must avoid both missing fake content and falsely accusing genuine footage of being manipulated.
The researchers also increased the system’s capacity to 18 simultaneous video streams. Its average accuracy remained at 96.13%.
These results suggest that optical computing could help process large volumes of video without requiring each stream to pass through a separate digital detection pipeline.
Researchers Test the System Against Google’s Veo 3 Videos
One challenge for deepfake detection technology is that AI-generated videos are becoming increasingly realistic.
Older detection systems often look for visual inconsistencies associated with early deepfake tools. Modern generative video models can produce more convincing footage, making these inconsistencies less reliable.
To investigate whether their technology could handle newer AI-generated content, the UCLA researchers tested it using videos created with Google’s Veo 3.
The system achieved 94.80% accuracy and 97.61% sensitivity on previously unseen Veo 3 videos after minimal fine-tuning.
This is an important result because it suggests the technology may be adaptable to newer video-generation models.
However, the experiment does not establish that the detector can identify every type of AI-generated video. Different generators, editing techniques, compression methods and recording conditions could affect its performance.
Why Using Light Could Make Deepfake Detection More Efficient
Conventional deepfake detection systems rely heavily on digital processors and neural networks.
As the number of videos increases, the computational resources required to analyse them can also increase.
The UCLA approach attempts to reduce this burden by moving part of the processing into an optical system.
Light propagation happens physically and can process multiple encoded streams in parallel. The passive optical decoder also does not require the same kind of active computation as a conventional digital neural network.
The researchers reported that their architecture could provide lower decoder energy requirements and reduced processing latency.
They also found that adding two optimised passive diffractive layers improved accuracy by approximately 6.8 percentage points in tests involving more difficult manipulations.
These layers perform additional processing through the diffraction of light, without requiring additional electrical power during inference.
The technology also showed resilience against several forms of adversarial manipulation and video degradation, including noise, blur and JPEG compression.
However, the complete system still uses digital components, including its initial encoder. It would therefore be inaccurate to describe the entire detector as requiring no electricity or digital computation.
Could This Technology Help Social Media Platforms Detect Fake Videos?
The researchers envision the optical processor as a first line of defence against large volumes of manipulated content.
Social media platforms, news organisations and content moderation companies receive enormous quantities of videos every day.
Running every video through a computationally intensive detection model can be expensive and time-consuming.
The UCLA system could initially screen multiple videos simultaneously. Content flagged as suspicious could then be passed to more advanced digital forensic models for deeper examination.
This creates a two-stage detection process:
| Stage | Purpose |
|---|---|
| Optical screening | Quickly examine multiple videos and flag suspicious content |
| Digital verification | Conduct detailed analysis of videos identified as potentially manipulated |
Such an approach could eventually support content moderation, media authentication, surveillance and other applications where large amounts of video need to be checked.
The researchers have not announced a commercial product or a confirmed deployment with major social media platforms.
The Bigger Picture
The research demonstrates how optical computing could contribute to the growing challenge of detecting AI-generated content.
As generative AI models become capable of producing realistic videos, detection systems will need to become faster, more efficient and more resistant to deliberate manipulation.
The UCLA approach is particularly interesting because it changes how part of the detection process is performed. Instead of relying entirely on conventional digital processors, it uses the physical properties of light to perform parallel computation.
However, high accuracy on a research dataset does not automatically translate into reliable performance across every real-world video. Independent testing on broader datasets and different generation techniques will be necessary before the technology can be trusted for large-scale deployment.
Looking Ahead
The researchers plan to continue developing the optical architecture, including improvements to its capacity and processing speed. Larger optical modulators and faster light-control components could allow the system to handle more video streams simultaneously while reducing latency.
The next major test will be whether the technology can maintain its accuracy against newer AI video generators and more sophisticated manipulation techniques outside controlled experiments. If it succeeds, optical-neural computing could become an important component of future systems designed to identify synthetic media at scale.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
