Deepfakes are video or audio content faked with the help of artificial intelligence (AI). Artificial Intelligence methods such as Machine Learning (ML) and Artificial Neural Networks (ANN) are used, for example, to exchange persons in video sequences or to have actors speak different texts. Some deepfake methods are applicable in real-time.
What is a Deepfake?
Deepfake is a word made up of the two words “deep learning” and “fake”. It stands for altering or faking video or audio content using Artificial Intelligence techniques. Software is trained with artificial neural networks and machine learning and is then able to replace faces or complete people in video sequences, change movements or have people speak different texts.
The altered content is difficult to recognize as fake. They appear genuine and authentic to the viewer. Constant improvements in deepfake software and increasing computing power ensure ever better quality deepfakes. Equipped with suitable software, even amateurs without specific video editing skills are able to create high-quality deepfakes. Deepfakes can be used for various purposes, such as entertainment, disinformation and propaganda, or to discredit individuals. Legally, deepfakes created without the authorization of the person concerned can violate personal rights or be considered insults.
How are deepfakes created?
Artificial neural networks and machine learning are used to create deepfakes. Artificial intelligence methods are used to analyze the existing audio or video material. This requires the source material to be faked and video or audio material of the people whose sequences are to be transmitted. The more material is available, the better the subsequent deepfakes.
For example, to swap faces in videos, also known as face swapping, footage of the people from different perspectives and with different facial expressions is necessary. After training, the software is able to transfer faces to other video sequences. The other contents of the video sequences remain unchanged.
There are numerous programs and apps that even laymen can use to create high-quality deepfakes. One very well-known application is the face-swapping app FakeApp. If the app is provided with image material of the person to be inserted into the source material, FakeApp generates the desired deepfakes almost independently using deep-learning algorithms and the AI framework TensorFlow.
What types of deepfakes are there?
The term deepfake covers the faking and altering of video content, images, and audio content. So-called face swapping is very popular in the area of faking images or videos. There are also deepfake types that exchange complete persons. This allows them to be presented in a completely different environment or context.
Another video deepfake type is body puppetry. In body puppetry, movements are transferred to other people. Voice Swapping is a deepfake type of audio content. Voice swapping allows texts to be spoken with the voices of strangers. Many of the deepfake types require relatively long preparation and training time before the actual fake is generated.
In the meantime, however, there are also methods that enable deepfakes in real-time. For example, a person’s lip movements and facial expressions can be captured and transferred in real-time to video images of other people.
For what purposes are deepfakes used?
Deepfakes are used for various legal or illegal purposes. Typical uses include:
- In the art and film industry
- For propaganda purposes
- For targeted disinformation or discrediting of persons
- To generate pornographic content
