Intel announces real-time deep fake detector 'FakeCatcher' using 'blood flow' in video, detection accuracy is 96%

As part of Intel's `` Responsible AI '' initiative, `` FakeCatcher '' that can detect fake images made with deep fakes with 96% accuracy has been announced.

Intel Introduces Real-Time Deepfake Detector

As part of Intel's Responsible #AI work, we've developed FakeCatcher, a new technology that detects fake videos with 96% accuracy. This detection platform is the world's first real-time #deepfake detector that returns results in milliseconds. More here. https ://t.co/jG2AakdFb8

— Intel Policy (@IntelPolicy) November 14, 2022

Intel unveils real-time deepfake detector, claims 96% accuracy rate | VentureBeat
https://venturebeat.com/ai/intel-unveils-real-time-deepfake-detector-claims-96-accuracy-rate/

FakeCatcher was co-designed by Ilke Demir, senior staff researcher at Intel Labs, and Umur Ciftci of the State University of New York at Bingham, and runs on servers using Intel software and hardware.


FakeCatcher uses OpenVINO to run AI models of face and feature detection algorithms on real videos. The computer vision block is the image processing library Intel IPP and the image processing toolkit OpenCV , and the inference block is optimized with Intel Deep Learning Boost .

While many deep-learning-based detectors look at raw data for signs of fraud and try to pinpoint issues, FakeCatcher detects the subtle “blood flow” in video that makes us human. By evaluating it, we will search for real clues. When the heart pumps blood, the veins change color. By collecting these blood flow signals from the entire face and converting them into a spatiotemporal map by the algorithm, deep learning can immediately detect whether the image is real or fake.

Potential uses of FakeCatcher include preventing harmful deepfake videos from being posted on SNS and preventing the inadvertent spread of deepfake videos by distributing them as news.

According to Intel, the 3rd generation Intel Xeon Scalable Processor can execute 72 deepfake detection tasks in parallel in real time, with a deepfake detection rate of 96%.