Currently being researched to make them more realistic, deepfakes – a contraction of deep learning and fake – allude to image synthesis techniques based on artificial intelligence. Modern, current deep-fake techniques are already yielding impressive results, which is a concern for US legislators.

There is a concern that they may be used against individuals to represent them in situations to which they would object, such as in pornographic videos. Indeed, it should be noted that this was the original purpose of this technology: it can be used in the porn industry, where faces can be replaced with others. According to US legislators, deepfakes also represent a real political danger. For example, they believe that this image synthesis technique could be used as part of broader misinformation campaigns to influence elections by broadcasting fake news that might then appear more credible. In other words, deep-fake creation technologies could be dangerous weapons in the hands of malicious people.

They use machine learning techniques called few-shot or one-shot learning, which allude to learning from a few examples or a single example. Indeed, if artificial intelligence is useful in applications that use a lot of data, it cannot learn from a limited number of examples. In other words, current deep learning methods suffer from low effectiveness for small samples, which contrasts sharply with human perception: even a child can recognize a giraffe, for example, after viewing a single picture. Few/one-shot learning, therefore, aims to provide deep neural networks with this ability: to achieve a high level of learning from small data sets.

In their work, Samsung researchers were able to use an image to create a convincing animated portrait. By increasing the number of shots (to 8 or 32 photos, for example), the moving picture becomes more and more realistic. And it even works on Mona Lisa’s portrait and other unique images. In a video published by one of the researchers on YouTube, the great pictures of Albert Einstein, Fyodor Dostoyevsky, and Marilyn Monroe come to life as if they were videos that were captured live from your smartphone’s camera.

The fact that Samsung’s AI needs only one photo to generate a living or deepfake portrait does not mean that the algorithm is based on only that one photo. Behind it, there is a preliminary meta-learning work that has been done. The researchers carried out extensive learning on a large database of a public repository of 7,000 celebrity images from YouTube. It allows the algorithm to identify what they call “historical” features of faces: eyes, mouth shapes, length, and shape of a nasal bridge.

From there, the system can generate a video or animated portrait from photos of people who were not included in the meta-learning phase. In concrete terms, when the system receives a new person’s photo as input, it initializes its parameters in a person-specific way, so that the generation of the deepfake or animated portrait can be based on a few images and can be done quickly, “despite the need to adjust tens of millions of parameters”.

As with most deepfakes, there are some flaws. Visual artifacts surround most faces. But the results remain impressive given that this technique is in its infancy.