Deep Fakes: Terror In A Data Driven World
What do recent viral videos of Mark Zuckerburg, Nancy Pelosi, and even the Mona Lisa have in common? They’re all examples of deep fake technology, an artificial intelligence technique that can generate incredibly realistic doctored videos.
Deep fakes are not new; they first appeared several years ago when anonymous users on Reddit posted doctored celebrity pornography. However, these early deep fake techniques required a large amount of photo data in order to generate accurate videos. Samsung’s AI Research Center has recently pioneered a new technique, capable of generating hyper realistic deep fake videos from just a single photo of the subject and an audio file. Using this technique, which relies upon GANs, a type of artificial neural network, a video is generated where the subject’s mouth moves according to the human speech in the audio file. Even small details, such as blinks, eyebrow movements, inflection, and emotion, are included.
The implications of this technology go beyond generating videos of the deceased Russian mystic Rasputin singing along to Beyoncé’s “Halo.” We are living in an increasingly data-driven world. More than ever, technology giants power their products with algorithms that rely upon massive streams of data. Furthermore, as competition increases, these streams of data are sought from increasingly obscure and oftentimes less reliable places—anything to get an edge. Deep fakes have the potential to “fool” algorithms that rely upon these data streams, making them less accurate and harming the companies that use them.
The risks of deep fake technology are not limited to corporations either. In the era of online media, deep fakes can be used to influence citizens and generate massive campaigns of disinformation. Indeed, we saw exactly this when a doctored video of Nancy Pelosi slurring her speech in a drunken manner, which was created using deep fake technology, went viral on social media. Deep fake technology can be used to generate outrageous but otherwise completely realistic videos of public figures, increasing the probability that they will be disseminated widely by an enthralled audience.
However, do not fear; just as AI has created this potential problem, it has also become the principal tool used to solve it. Current deep fake videos often involve temporal inconsistencies—small, awkward movements that would not occur with natural speech. Researchers have created a machine learning model capable of using these inconsistencies to identify deep fakes with over 90% accuracy. By implementing these models into their data pipelines, it seems that companies can largely mitigate the risk of deep fakes. That said, this will certainly be an arms race. As deep fakes become more realistic, we will have to turn to increasingly sophisticated techniques to detect them.
Written by Daniel DiPietro, Edited by Matthew Durborow & Alexander Fleiss
We just sent you an email. Please click the link in the email to confirm your subscription!