A deepfake detector developed at the University of Colorado Boulder starts with the information already carried in compressed video. By screening those clues before reconstructing full images, DeepSFU used about one-143rd of the computing power of the existing detectors in the team’s comparison.[1]
Presented at ACM SIGCOMM, the system also achieved what the researchers describe as near-perfect accuracy on a collection of 1,000 online videos containing deepfakes. The team wants to adapt it for large video conferences, where continuously examining every participant’s full-resolution stream would be expensive.[1]
What an ordinary video compressor already knows
Video compression records changes between frames. A static background needs little new information; a moving face needs more detail. DeepSFU measures how much new information falls on the face and how widely it is distributed, looking for a suspicious pattern in the encoded stream.[1]
The initial screen reads data from roughly one in every hundred frames. When a pattern looks suspicious, it requests a fresh, complete frame from the sending device for a visual AI check. The design concentrates the more expensive image analysis on selected moments.[1]
How DeepSFU allocates the work
| Stage | What the system examines | Purpose |
|---|---|---|
| Initial screen | Patterns in compressed video | Identify a possible anomaly cheaply |
| Second look | A requested complete frame | Run visual deepfake analysis |
| Reported evaluation | 1,000 collected online videos | Compare accuracy and computing requirements |
Why the network pipeline matters
The saving comes from choosing where to inspect the video. A conferencing service already receives an encoded stream; using its information can reduce repeated decoding and visual inference. The benefit grows with the number of simultaneous streams that a service must handle.[1]
Live deployment would need to test different codecs, bit rates, cameras and network conditions. Changes in the encoding pipeline can change the clues available to a detector. False alarms also need measurement over ordinary calls, where genuine users greatly outnumber impersonation attempts.[1]
The request for a fresh frame creates a separate security question: what assurance does the service have about the responding device? A malicious sender could control its output. That is a deployment question raised by the architecture, rather than a failure demonstrated by the collected-video result.[1]
A cheaper first screen
DeepSFU shows that detection can begin before a compressed stream becomes a picture. Its next practical test is whether that inexpensive first screen retains useful accuracy and a manageable false-alarm rate across live conferencing conditions.[1]
Sources & further reading
- AI deepfakes are getting more realistic. CU scientists developed a better way to spot them University of Colorado Boulder · Accessed 25 Oct 2026
- DeepSFU: Scalable Deepfake Detection for Video Conferencing ACM SIGCOMM conference paper · Accessed 24 Oct 2026