Imagine looking into a mirror and not seeing your own face. Instead, you are looking in the eyes of Barack Obama or Angela Merkel. Your facial expressions are seamlessly transferred to that other person's face in real time.
The TNG Hardware Hacking Team has managed to create a prototype that transfers faces from one person to another in real time, based on Deep Fakes. Neural networks detect faces within a video input, then translate them and integrate them back to the video output. Through this technique, it is possible to project deceptively real imitations onto other people. For this purpose, we used Keras trained autoencoder networks and various face recognition algorithms. In the talk, Thomas Endres and Martin Förtsch will give an entertaining and vivid introduction into the world of real-time Deep Fakes. In doing so, they particularly focus on the deep learning techniques used in within the application. Several live demonstrations will complete this event.
Keywords: Tensorflow, OpenCV, Keras, UNet, MobileNet, Transfer Learning, Deep Learning, Autoencoder, Realtime Deepfake