Deepfake (a portmanteau of deep learning and fake) is the synthesis of fake images, video and sound using artificial intelligence. As a countable noun, a deepfake is a piece of content obtained through such synthesis. Most often, deepfakes depict famous people in made-up situations. In some cases, deepfakes also include AI-generated written text to imitate the speech of the target.
Modern deepfake tools can generate photos, videos, audio and text in real time (for example, in the case of a live video broadcast).
How deepfakes are created
Deepfakes are created using deep machine-learning technologies.
Audio and graphic deepfakes are most commonly created using a generative adversarial neural network (GAN). Such a network includes two algorithms: a generator and a discriminator. The generator creates an image or sound, while the discriminator, trained on a large array of data, determines the credibility of this image or sound. In doing so, the system learns from itself: the generator uses feedback from the discriminator to generate increasingly convincing content; the discriminator utilizes the result of the generator to better distinguish original from AI-generated data.
Text deepfakes are typically created using transformers — neural networks capable of processing and generating datasets consisting of separate yet interconnected segments (such as words, sentences and paragraphs). Such neural networks generate text based on the probability of each next element in the sequence occurring.
Legitimate use of deepfakes
Deepfakes can be used for legitimate purposes, such as:
- Reducing the budget and production time of movies, TV programs, commercials, music videos, etc.
- “Reviving” a celebrity from the past for a media project.
The danger of deepfakes
Deepfakes are often made and used for criminal purposes. In particular, deepfakes are associated with the following risks:
- Attackers can create and distribute content in which public figures appear to say unacceptable things, act under the influence of alcohol, or engage in activity of an inappropriate, sexual and/or criminal nature.
- Deepfakes can be used to spread fake news and influence public opinion.
- Financial fraud. By imitating somebody else’s voice or writing style using AI, a cybercriminal could pose as a top manager and, for example, ask a subordinate to transfer money to their accounts.
- Disclosure of personal data. Scammers can find out personal information about the target from their friends or relatives through audio or video communication with them.
- Fooling authentication systems. Fraudsters can pass audio and video authentication to gain access to money, valuable data or corporate information systems.