An introductory example using a fake Obama video highlights the potential for misinformation and the importance of public awareness.
A breakdown of related but distinct technologies like lip syncing, face swapping, and real-time reenactment clarifies what constitutes a genuine deepfake.
The original deepfake workflow involves preparing a dataset with FaceNet, training a shared autoencoder, and then running inference to generate the final video.
To achieve real-time performance, the original pipeline is modified with faster face segmentation using MobileNet and U-Net, and image inpainting to remove the original head.
Generative Adversarial Networks (GANs) are introduced to the training process, where a generator and a discriminator compete to produce more realistic and artifact-free faces.
A live demonstration shows the real-time head replacement and reveals a key limitation where the model can only reproduce expressions seen in its training data.
Deepfakes offer a cost-effective alternative to CGI in the movie industry and can also be used to train discriminator models for detecting fake news.
The next evolution of deepfake technology will likely involve higher resolutions, full-body manipulation, and real-time voice cloning to create more comprehensive fakes.
30:02Martin Förtsch & Thomas Endres
30:38Timo Zander
56:55Ekaterina Sirazitdinova
24:49George Proorocu
21:01Jan Schweiger
47:28Tanmay Bakshi
32:54Adrian Spataru & Bohdan Andrusyak
29:54Fabian Pottbäcker, Thomas Endres & Martin Foertsch
.gif?w=240&auto=compress,format)
Jobs that call for the skills explored in this talk.
Technische Universität Berlin
Deutsches Medizinrechenzentrum GmbH
DTS Systeme GmbH
Fraunhofer-Gesellschaft zur Förderung der angewandten Forschung e.V.
Berlin, Germany
ImFusion GmbH
München, Germany
NeXaT GmbH
Bersenbrück, Germany
Forschungszentrum Jülich GmbH
Jülich, Germany