False detection by blood flow in video
As part of Responsible AI, or responsible AI, Intel announced FakeCatcher which can detect fake videos made with deep fakes with 96% accuracy.
FakeCatcher runs on servers using Intel software and hardware co-designed by Intel Labs senior researcher Ilke Demir and State University of New York Umur Ciftci. It uses OpenVINO on real video to run AI models of faces and specific detection algorithms. The computer vision block is optimized with the Intel IPP image processing library and the OpenCV image processing toolkit, and the inference block is optimized with Intel Deep Learning Boost.
Many deep learning detectors examine raw data to find signs of malignancy and try to identify problems, but FakeCapture evaluates subtle blood flow in images to find real clues. When the heart pumps blood, the color of the veins changes. By collecting these blood flow signals from the entire face and converting them into space-time maps through an algorithm, deep learning is said to be able to instantly detect whether the image is real or fake.
Examples of possible uses of fake capture include preventing SNS from posting harmful fake videos or carelessly spreading fake fake videos presented as news. According to Intel, 72 deep fake detection tasks can be performed in parallel in real time on a 3rd generation Intel Xeon Scalable processor, and the deepfake detection rate reaches 96%. Relevant information can be found here.
