F for Fake

F for Fake

F for Fake. No, not the 1973 Orson Welles movie by the same name, but rather deepfakes – one of the scarier AI derivative technologies.

Samsung’s “Few-Shot Adversarial Learning” models could render a video of you from a single profile photo. Researchers at Samsung’s Moscow-based AI research center have developed a “deep convolutional network” capable of producing a deepfake video from a single profile pic. Deepfake video quality improves as additional pics are introduced to the model.

This gets a bit scary when you consider the implications of weaponizing deepfake videos. From discrediting and/or persecuting political opponents, to discrediting CEOs whom activists may wish to expunge from the companies they run, to various fraud-related use cases. The latter example is a very real threat, particularly as we move to facial recognition as a source of payment authentication and digital identity authentication.

Samsung’s research effort was presented in a paper called “Few-Shot Adversarial Learning of Realistic Neural Talking Head Models” which may be accessed HERE.

Bringing the Mona Lisa to life with AI

Minimal learning required. Samsung AI Research Center. Research paper by: Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, Victor Lempitsky. https://arxiv.org/abs/1905.08233v1

Google Deep Learning depth map