1. What are deepfakes?
  2. How are they made?
  3. How are they used now?
  4. What are their potential benefits?
  5. Which dangers do they pose?
  6. So, how exactly do scammers use deepfakes?
  7. НIs it really that hard to recognize a deepfake?
  8. Are there any other ways to detect and counteract deepfakes?

What are deepfakes?

A deepfake is a modified bit of photo, audio, or video content made via a neural network. The word itself is a combination of “deep learning” (a type of machine learning – Ed.) and “fake”. Such algorithms are capable of making still images move, synthesizing voices, and swapping faces in videos. With all of these tools, users can generate content that looks almost indistinguishable from the real thing. This technology is relatively new, as it was developed in 2014 by Ian Goodfellow, a student from Stanford University.

How are they made?

They are created using generative adversarial networks (GAN). These are machine learning frameworks that can make new content by using a training set of data. For example, by showing the neural network several thousand pictures of Mr. Bean, you can then create a separate character with the same facial features and expressions. Or superimpose his face on top of someone else’s – just like in this fake Dior ad starring Mr. Bean. Such networks are called adversarial because they are pitted against each other – the first creates a video, the second tries to verify its authenticity. If the latter spots a fake, the former gets feedback on what to fix and starts over. This process is repeated until the first network produces something that the second network finds convincing.

How are they used now?

They have different uses. Entertainment is the least dangerous. For instance, in 2020, a deepfake video of Elon Musk performing the classic Russian song Trava u doma (Grass by the Home) went viral. A year later, a series of videos emerged of Tom Cruise doing magic tricks, playing golf, and speaking of his meeting with Mikhail Gorbachev, among other things. The actor Miles Fisher played Cruise, his face replaced by a realistic digital mask that was created using a neural network. The success of these videos led their creator, the Belgian visual effects artist Chris Ume, to launch his own studio specializing in deepfake technologies.

In 2022, an entire web series named PMJason was produced in Russia using deepfakes. The show takes place in 2027. Having wrapped a five-year shoot in Russia, Jason Statham decides to stay there and settle in a village. On his 60th birthday party, he is visited by his friends Keanu Reeves and Margot Robbie.

What are their potential benefits?

Some companies use deepfakes to produce their corporate training courses. For instance, the 50,000 employees of the British advertising giant WPP are learning the basics of marketing via videos with a virtual teacher who can speak three languages: English, Spanish, and Chinese, while addressing each of its pupils by name. The course cost $100,000 to make, but the company claims that using live actors would be 10 times more expensive. Apart from WPP, Reuters, SAP, Accenture, and Buzzfeed also utilize training videos made by neural networks. These materials are used to teach newcomers about occupational health and safety rules.

Which dangers do they pose?

This technology can be used for fraud and manipulation. There have already been a lot of such cases. Last year, two scammers from China managed to trick the tax service’s facial recognition systems and make off with $76.2 million, while in Russia, deepfake ads were created using the likeness of businessman Oleg Tinkov. The artificial billionaire in the video urges people to invest and receive prizes by registering via a link. In the UAE, scammers used deepfake to clone the voice of a CEO of a major company and deceive bank managers into transferring $35 million to their account. Computer security experts think that such scams are going to become even more prevalent.

So, how exactly do scammers use deepfakes?

Nowadays deepfakes are mostly used to discredit individuals. For example, there are cases from all over the world when they were used to damage the reputations of journalists. This is especially true for India, where technological literacy is not at a very high level.

Is it really that hard to recognize a deepfake? It seems like all it takes is simply paying attention to facial expressions and voice inflections.

Lately, high-quality deepfakes have become quite hard to distinguish with the naked eye. If the photo or video was also processed in some way, then it is almost impossible. A couple of years ago deepfakes could not mimic blinking, since they used static images to train. Nowadays this flaw has been eliminated. In general, the signs of a deepfake are: a disproportionately large or small head, unnatural lighting, wrong placement of shadows, and blurry details like eyelashes or eyebrows (the Oleg Tinkov deepfake is a good example).

Are there any other ways to detect and counteract deepfakes?

There are numerous methods: a few involve studying the video frame-by-frame, some look at the skin texture, others search for artifact errors, etc. In general, it is easy for an expert to detect a deepfake, with some technologies for verifying the authenticity of digital content already available to the public. For example, in 2022 Intel released FakeCatcher, an artificial intelligence framework that can spot deepfakes within videos in real time. The developers claim that their system can distinguish a real person from a fake by, among other means, examining the change in color of their blood vessels. The algorithm collects that information by observing multiple spots on the face, and then uses AI to process the data.

It should be noted that Ilya Popov, the head of the Laboratory of Validation and Verification of Complex Technical Systems of ITMO’s Faculty of Secure Information Technologies, and Vitaliy Rogovoy, a research associate at the laboratory, also spoke of their work on deepfakes and their identification at the AntiFraud Russia 2022 forum, which took place in December.