Four things you can do to fight against deepfakes

Earlier this year, the FBI released a warning stating that malicious actors “almost certainly will leverage synthetic content for cyber and foreign influence operations in the next 12 to 18 months.” Most of us know synthetic content under the term “deepfake” — a blend of the words “deep learning” and “fake.” It is a form of artificial intelligence (AI)-generated synthetic media that first emerged on internet forums in late 2017 and have since found their place predominantly in entertainment.

Nowadays, you’ll mostly come into contact with deepfakes created and shared for comedic purposes — many of us have, at some point, downloaded one of the many apps on the market that swap the face of our favourite movie stars or superheroes for our own. And even though the results can sometimes be eerily realistic, we share them light-heartedly with our friends.

Although some deepfakes are quite easily spotted and are not malicious in intent, the technology is advancing to the point where some deepfakes cannot be discerned from authentic videos with the naked eye. Currently, users are more likely to be deceived by information whose context has merely been altered — in other words, fake news — rather than information that has been completely synthesised, the FBI report reads.

This, however, is bound to change as AI and machine learning (ML) technologies become more sophisticated and convenient.

Four troublesome heads

Deepfakes are essentially a new type of fake news — content created to misinform or deceive readers deliberately. Fake news is nothing new; it has always been around. The main difference is that now, owing to the speed at which AI-based tools can encourage and facilitate the sharing of information online, false stories can spread faster, at a larger scale and consequently do more harm.

Similarly to fake news, the actual manipulation of images and videos has been around for decades: for example, Stalin notoriously had his political enemies erased from photos. There are examples of special effects used in video dating back a century, such as the famous silent film from 1898 by Georges Méliès, The Four Troublesome Heads, where he used a multiple exposure of objects on a black background to create the impression of detached heads.

Today, you have the tools to create deepfakes right in your pocket. A wide range of free apps is available on the market that allows users to superimpose celebrities and movie characters. Although most of us could spot the 1898 special effects as easily as recognising our friend in a Maleficent Halloween costume, professional deepfakes today are made to be as convincing as possible — at times very successfully. For example, a number of videos emerged this year impersonating the American actor Tom Cruise, and the results are impeccable.

The emerging world of deepfakes

What is the technology behind this emerging trend? Companies working with deepfakes can utilise a type of ML method, a subset of AI, called Generative Adversarial Networks, or GANs. The method sets two neural networks in competition with each other, where one is the generator and the other the discriminator. In other words, one neural network learns to generate deepfakes, while its counterpart learns to discern whether the image is real or fake — like a game of cops and robbers where both sets continuously improve themselves against the other.

“After millions of attempts, the discriminator becomes better and better at recognising deepfakes and is able to reach unexpected results,” shares Martin Černý, head of the global partner network at Cogniware, a Czech company that has been developing specialised software called Insights for national security forces to assist in criminal investigations of fake news. Their focus is text-based misinformation, but they have recently worked on a bespoke project to develop a solution for identifying deepfake videos.