Deepfake, what they are and how they are created

Deepfake

With deepfake we mean the practice that allows you to create fictitious images, photos or videos, but using elements of real photos or videos. In practice, through this technology it is possible to replace the face of one person at will with that of another, trying to obtain a result that is as realistic as possible. Furthermore, it should be noted that, which is easily predictable, to date 96% of the deepfakes existing on the Internet concern the world of porn and see the replacement of the face of the original pornstars often with that of famous actresses, while the remaining 4% concerns situations mostly. comics.

How are deepfakes created?

The peculiarity of deepfakes is that they manage to do all this in a convincing way. There are essentially three steps required to create a deepfake. Initially, the required information needs to be retrieved, such as the two faces you want to exchange. Next, we need to train the network to learn how to replace the faces and, finally, apply the results of the calculations to the final movie.

In the first phase, a software has to extract from a video every single frame of face A, going to get hundreds of frames in which that face takes on different expressions and frames. The same must be done with face B, since the final goal is to try to match the two faces with the same expressions and shots.

In the training phase all these photos are fed to the software who "extrapolates" the recurring information and with them creates new images, which however are not copies of the original ones, but completely "false" and which have the purpose of being realistic. In practice, in this phase the software learns to create images that represent real facial expressions, using the data that has been supplied to it.| ); }

What are the neural networks used for deepfakes?

Up to now we have talked generically about software to simplify concepts to the maximum and to better understand the approach of how deepfakes are made. Going into more detail, this technique can be achieved thanks to the training of what are called "neural networks", that is, complex algorithms that require a lot of power and that are made in such a way as to learn, which in this case means "refine their results proceeding by attempts ".


A second network, defined as a "discriminator", has the task of analyzing the image created by the generative network and saying whether it is a real image or whether it is the result of an artificial reconstruction. This network analyzes whether there are defects in the image and the two antagonistic networks challenge each other, in a certain sense, until the generative network is able to create images that the second network will always recognize, or almost always, as real. At that point we are faced with an infallible AI capable of creating super realistic deepfakes.

It must be said that today's deepfakes are not at this level and they are not all based on GAN networks, indeed often others are used algorithms which, in any case, perform the three steps indicated initially. But the final result, that is the photo or the deepfake video, also goes through various editing stages that allow to mask the defects that the neural network has not been able to correct. In addition to photo or video deepfakes, there are also audio deepfakes that allow, for example, to analyze and sample your voice and create invented audio recordings.




Audio deepfakes are more problematic today, as they are more difficult to identify and can be used for scams. Imagine getting a call from your boss asking you to hand him some data; your reaction will probably be to do what he asks you and even if the voice is not perfect, you would probably blame the sound of the call being a little disturbed. Such episodes have already happened and cybercrimes have been carried out in this way.

In short, deepfakes can be potentially harmful and if today they are mainly perceived as occasions for some laughter and solace, tomorrow they could become a risk, both personal and professional.