Deepfakes are moving from viral clips to enterprise verification, where camera feeds serve as proof for onboarding, account recovery, and privileged access. Purdue’s PDID benchmark tests detectors on real-world, messy social-content, revealing Deepsight’s production-ready performance and a layered defense that protects media and decision paths from capture to verification. #PDID #Deepsight #IncodeTechnologies #PurdueUniversity #VirtualCameras
Keypoints
- Purdue’s PDID benchmark uses real-world, messy social-content (heavy compression, sub-720p, short clips) to test detectors.
- PDID includes 232 images and 173 videos, evaluated with accuracy, AUC, and FAR to reflect production conditions.
- Deepsight achieves the lowest image FAR of 2.56% with 91.07% accuracy and the best commercial video accuracy of 77.27% with 10.53% FAR.
- Deepsight uses a three-layer defense—Perception, Integrity, and Behavioral—to protect media and decision paths in real time.
- In real-world deployment, Deepsight reduced false-acceptance by 68x, identified 10x more deepfakes than humans, and caught 24,360 fraudulent sessions.
Read More: https://thehackernews.com/expert-insights/2025/12/purdue-universitys-real-world-deepfake.html