New technique uses light to spot deepfakes in seconds

In a breakthrough in the fight against AI-generated misinformation, researchers at the University of California, Los Angeles (UCLA) have developed a system that uses light to detect deepfake videos in seconds.

According to findings published in the journal eLight, the technique could significantly reduce the burden on social media platforms, which currently rely on resource-intensive electronic systems to process large quantities of user-submitted videos. Recent research in Nature Communications found that humans can spot deepfakes only about 50% of the time.

The technique combines an electronic computer with an optical computer — a machine where light performs part of the work normally done by chips.

The system works by selecting 12 frames from a video and checking the faces in them for patterns. It then feeds the gathered information into a device called a spatial light modulator and then to two spots on a sensor. Based on the difference in brightness between these spots, the system determines whether a video is fake.

In lab tests using the Celeb-DF database of face-swapped videos, the system identified manipulated videos with 99.86% accuracy and real videos with 95.72% accuracy, while processing 15 videos at a time.

The researchers said the system also identified videos made with Google’s Veo 3 AI model with 94.8% accuracy, showing it does not need to be retrained from scratch against the latest AI models.

Crucially, the optical part of the machine consumed seven times less energy than the electronic part, making it a low-energy alternative for large-scale screening.

The team now plans to scale up the architecture using optical modulators with larger aperture to multiplex substantially more videos and faster modulators to further reduce latency.