Deepfakes and the world of Generative Adversarial Networks
--
GANs, short for Generative Adversarial Networks, were introduced in a paper by Ian Goodfellow and other researchers at the University of Montreal, including Yoshua Bengio, in 2014: We propose a new framework for estimating generative models via an adversarial process, in which we simultaneously train two models: a generative model G that captures the data distribution, and a discriminative model D that estimates the probability that a sample came from the training data rather than G. The training procedure for G is to maximize the probability of D making a mistake. This framework corresponds to a minimax two-player game.
MIT IBM Watson AI Lab has released a Gan Paint Demo that gives a good insight into how the GAN technology works.
GANs can be used for a number of exciting things but what has caught the public’s imagination is the use of GANs to create deepFakes, i.e. to create videos of talking people where the face has been swapped for some else. Such as the video of Barack Obama above.
The deepFake space consists of numerous cloud-based webpages such as https://deepfakesweb.com and the Chinese app Zao, that allow anyone with a minimum knowledge of the web to create deepFakes and host them in the cloud.
So GANs can be used to create deepFakes. But deepFakes is only a very small part of the story of GANs.
GANs can be used to create photos of imaginary fashion models, with no need to hire a real model, photographer, makeup artist, or pay for a studio and transportation.
Generate fashion images using Generative Adversarial Networks. Define a GAN model and implement a DCGAN model to generate fashion images without writing a single line of code.This IBM article, by Smruthi Raj Mohan, Raunak Sinha, Anush Sankaran and Senthil Mani, describes a code pattern that is made up of a simple JSON that represents a Generative Adversarial Network (GAN) model that implements a Deep Convolutional GAN (DCGAN) model to generate fashion images without writing a single line of code.
And the data set that makes this possible in the fashion industry is
Fashion-MNIST, a dataset of Zalando’s article images — consisting of a training set of 60,000 examples and a test set of 10,000 examples. Each example is a 28x28 grayscale image, associated with a label from 10 classes.
This is an example of legitimate uses of GAN models, that can be quite dramatic, replacing humans in an entire industrial segment.
GANs can also be used for creating artificial medical images for testing purposes, such as Generating MRI Images of Brain Tumors with GANs, which Alisya Kainth has described in Medium and GANs for creating artificial medical images, described by Edwin Donelly on Twitter.
Using GAN for Medical Imaging is a rapidly growing field primarily because of the existence of large datasets.
Please note the difference between GAN and DCGAN, Deep Convolutional Generative Adversarial Networks.
So we see how GAN-based apps are depending on a number of factors. First of all the underlying neural network technology, then large datasets and then in the case of deepFakes, easy-to-use client tools on the web that allows anyone with a laptop to quickly create deepFakes.
A very interesting and at the same time worrysome technology.
