A team of researchers at the University of Colorado Boulder has developed a system that spots deepfakes in video calls using 143 times less computing power than existing detection models. Tested on a library of 1,000 online videos containing deepfakes, the system reported near-perfect accuracy.

Instead of reconstructing every frame of every participant’s video at full resolution, the system reads clues directly from the compressed video that travels over the network, and only asks the sender’s device for a complete, fresh frame to check with artificial intelligence when something looks suspicious. The team presented the work, DeepSFU: Scalable Deepfake Detection for Video Conferencing, at SIGCOMM, a flagship conference in computer networking and systems held in August in Denver.

  More: Longmont Leader
 

The University of Colorado Connection

The research team was co-led by Sangtae Ha , associate professor of computer science at CU Boulder, and Eric Keller, professor of computer engineering at CU Boulder. Tuan Tran , PhD candidate of computer science at CU Boulder, and Shirin Ebadi, PhD candidate of electrical and computer engineering at CU Boulder, are co-authors of the paper. Keller said the team wants to adapt the system to scale up for large-scale video conferencing.

Both co-leaders have ties to CU Boulder’s startup community. Ha is an advisor to Choyu Networks Inc., a startup formed to bring post-quantum cryptography for low-power devices and limited-bandwidth networks to market and a 2026 finalist in CU Boulder’s Lab Venture Challenge (LVC). Keller previously co-founded Stateless, a startup built on network virtualization technology developed by Keller and then-PhD student Murad Kablan, which was selected for the 2017 Techstars Boulder accelerator.

More: CU Boulder Today