Computer scientists at the University of Colorado Boulder have created a new deepfake detection system that may soon help protect people during live video calls. This technology can spot AI manipulation in real time and uses much less computing power than current security tools.
At a recent computer networking conference in Denver, doctoral candidate Tuan Tran showed why this tool is needed. Half the audience could not tell the difference between an AI-generated video of Brad Pitt and the real actor. Cybercriminals are now using these advanced models to impersonate company leaders. In one recent case in Hong Kong, an employee was tricked into sending 25 million USD to fraudsters.
Even with these rising financial risks, big platforms like Zoom and Google Meet do not yet have live deepfake monitoring. Existing detection tools are too costly and use too many resources because they have to decompress and check every frame of a live video at full resolution.
The CU Boulder team, led by professors Sangtae Ha and Eric Keller with doctoral candidate Shirin Ebadi, found a way around this problem by looking at how video moves across the internet. Rather than rebuilding the video, their system reads the compressed data directly.
Video compression saves data by recording changes in pixels between frames instead of sending the same background over and over. The researchers found they could detect fake changes just by measuring how much new pixel data appears around a caller’s face.
With this simpler method, a high-resolution AI check only happens if the compressed data looks suspicious. By checking just one out of every 100 frames, the team cut the needed computing power by 143 times.
When tested on 1,000 fake videos, the tool was nearly perfect. The team is now working to expand the system for use in large company video networks to fight the growing problem of digital impostors.