The term ‘Fake News’ has recently been named word of the year by Collins Dictionary with a 365% increase in popularity. While Trump uses the term to describe the US news networks, the phrase is has now entered the mainstream worldview as reality.
It is defined in the newest edition of Collins Dictionary as “false, often sensational, information disseminated under the guise of news reporting”. Essentially it is modern day propaganda.
The news could be be getting more fake
I want to bring to light some recent and not so recent developments in the fields of AI, image manipulation and rendering that really can create convincing material from thin air, generating audio, video and realistic faces.
Real-time Generated, Fake Video
Researchers at The Visual Computing Lab at Technical University Munich have many projects ranging from real-time scanning, recognition and generation of 3D environments, augmented reality and video generation. Perhaps the most disturbing is their Face2Face, a real-time capture and reenactment project that can take footage from a live actor and transfer it onto a live video.
Deleted Reality
Another research project by Jan Herling and Wolfgang Broll, Ilmenau University of Technology, Department of Virtual Worlds / Digital Games shows how even real objects can be removed in real-time. They are calling it Diminished Reality.
Tech companies continue to push the boundaries between the real and the computer generated
Nvidia are using Deep Learning AI that not only drive cars, but also run cities using camera feeds, security networks and deep learning AI to identify and analyse objects, people and faces from video feeds. Remind you of Orwell’s 1984 anyone?
Fake Faces
Nvidia have also used AI to capture facial performance and transfer the generated animation onto a virtual character. They have also recently released a video where they show completely fabricated faces made by the AI in an itterative process. Using Deep Learning and GAN methods they have managed to cross the uncanny valley into very realistic territories.
Fake Speech
Google DeepMind is in development of a generative model called WaveNet that can synthesize very human-sounding speech that is a vast improvement on standatd Text-to-Speech algorithms. Taking in information about inflection from the source audio sample along with phonetic data from the text, WaveNet generates new audio.
How do the AI learn?
Generative Adversarial Networks (GAN) train AI through input of large amounts of reference data, be it text, imagery or audio to generate new data. One AI (The Generator) generates fake data, which, along with real data, is fed into another AI (The Discriminator), that determines if the data it is receiving is real or fake. Together, the AI train each other, The Discriminator becoming better at distinguishing real from fake and The Generator AI getting better at creating realistic data.
With the rise of AR and AI combined technologies and all the recent developments it’s not a small stretch of the imagination to see an augmented world where the reality before us is edited and manipulated, where fake celebrities tell fake stories on fake news channels, maybe we’re already there…
The video below summarises the recent developments and points towards more sinister applications.
It is said that we are living in a 'Post Truth' era, but I don't think it needs to get to that stage. By critically analysing your sources there are still methods that can distinguish real and fake for now, but with the rise of Deep Learning AI and generated media, this might change in the future. All we can do is get our news from as many sources as possible, this way we can have more information to compare against and make better 'real or fake' decisions ourselves.
What do you think about Fake News, the rise of AI and image manipulation technologies? Do you have any thoughts on where these technologies could be applied for the greater good? Perhaps you working on a GAN project of your own?
Let me know in the comments below :)
