Generative Adversarial Networks have the potential to generate quality content out of noise. Ths capability can be put to good use for various tasks. However, at times the same capability can be misused to spread fake/false information. DeepFake videos are a classic example of this misuse. Faces/audio in one image/video is swapped with other image/video and the swap is made to appear as real as possible!
Our(VJTI,Mumbai's AI & Deep learning team) attempt is to demonstrate the technology on a lighter side and spread awareness. From technology perspective we built an architecture based on Generative Adversarial Network (GAN) and Autoencoders to swap face of two persons in the video. For purpose of demonstration we have swapped JackMa on AamirKhan in iconic 3 Idiots movie. The models where trained on NVIDIA V100 GPU. Watch carefully how eyes and lips are synchronized in the video. https://www.youtube.com/watch?v=f1zh0dQ3nrA&t=1s
The architecture and losses developed will be published in the research paper. The main aim of the research project is to detect DeepFake videos in the wild. Inorder to make this possible we are analyzing DeepFake generating methods inorder to exploit their irregularities.
The AI & DL team photo working at CoE-CNDS VJTI, Mumbai:
The page will be updated with ongoing research work.