Fakes at the touch of a button

What is a deepfake, and how is it created?

The term derives from the process of generation: “Deep learning” with images, videos or audios is used to train artificial neural networks and generate a fake that is as perfect as possible – i.e. a deepfake.

A very common method is the use of Generative Adversarial Networks (GANs). In this method, two networks are pitted against each other: One – known as the “generator” – is trained to create fakes that are as realistic as possible. The other – the “discriminator” – is designed to detect and evaluate these fakes. “As these two networks are played off against one another, both the counterfeiting and the control entity become more and more accurate, which means that the result gets more and more realistic,” explains Vasilios Danos, AI expert at TÜVIT.

This is a highly complex process, but one that can be carried out at the touch of a button. All you need is appropriate software and sufficient source material to train the AI. And video or audio recordings of celebrities from the worlds of politics, business and culture are everywhere on the Internet and thus easy to find.

There are now a variety of websites and smartphone apps that make the creation process even easier. “It's never been easier to create realistic fakes and then distribute them on social media,” says Mr. Danos. “And by the time it’s clear that it’s a deepfake, the damage has usually already been done.”

The larger the image or the higher the resolution, the easier it is to detect inconsistencies and flaws in the image. So, if you have doubts about a dubious video, you should watch it on a larger monitor and not on your smartphone. Pausing the video or playing it in slow motion can also reveal errors, thereby exposing manipulation.

As with other sensational content on the internet, the following applies in this case too: Healthy caution is often a wise move. The key questions to ask yourself is this: Can this content also be found on reputable news sites? Have the images or videos already been reviewed by fact-checking portals such as Mimikama or Correctiv? And, above all, are the statements and behaviour of the person shown consistent with what he or she usually says and does? “If, for example, German Chancellor Olaf Scholz appears to say or do something scandalous, your first response should be one of scepticism,” says AI expert Mr. Danos. “Especially when you realise how easy it is to create and distribute deepfakes.”

In many cases, AI is not used to generate a completely new video but only to replace a person’s face or the area around the lips to put other words in the speaker’s mouth. For this reason, in suspect pictures or videos, you should pay attention to the “seams”, i.e. the transitions between the face and the hair or the face and the neck. These are often still blurry in deepfakes. If the face has a different quality than the rest of the image or video, and the shadows, lighting or movement around and of the mouth, say, do not match the rest of the face, this is also suggestive of a deepfake.

When we speak or have conversations, our face executes a complex series of motions: We blink and frown; and, when things get heated, that pulsing vein of anger can appear. AI is not yet able to mimic such subtleties of human facial expressions that well. A closer look, especially when the image is slowed down, can reveal when an image is a fake. In the case of photo deepfakes, looking at the hands and fingers can be revealing. The generators often fail to accurately reproduce these details, leading to anomalies. For example, one hand may suddenly have six fingers, or arms may merge with other objects.