Deepfake videos are getting harder to spot by eye. A team at UCLA has built a detector that does much of its work with light rather than electronics – and it can check many videos at the same time.
How it works, in plain terms
- A small conventional computer program pulls out the key features of each video.
- Those features are turned into a pattern of light using a programmable light modulator.
- The light passes through specially designed layers that need no electrical power to do their part of the calculation.
- Light detectors at the end produce a simple score: real or fake.
The numbers
- 97.79% average accuracy when checking 15 videos at once
- 96.13% accuracy with 18 videos in a single pass
- 94.8% accuracy on videos made with Google’s Veo 3 video generator
It was tested on a well-known research collection of celebrity deepfakes (Celeb-DF) as well as Veo 3 videos.
Why it matters
Checking video for fakes at scale takes a lot of computing power. Because the optical part runs on light and uses no electricity during each check, this approach could make large-scale screening faster and cheaper.
The catch
The researchers describe it as a first-stage screening tool, not a final verdict. It also needs retraining as new video generators appear – which is the never-ending problem with all deepfake detectors.
Want to sharpen your own eye? See our guide: How to spot a deepfake video: 8 things to check.
More on Contoh
- How to Spot a Deepfake Video: 8 Things to Check
- Why AI Chatbots Make Things Up — and How to Catch Them When They Do
- An AI Just Crushed the World’s Best Stratego Player — on a Budget of a Few Thousand Dollars