Facebook announces specific technology of AI model that created ultra-realistic AI fake image 'Deepfake'

'Deepfake, ' which synthesizes fake human images using artificial intelligence (AI), is a technology that is easily used for fake news and is also regarded as a problem because of the damage caused by revenge pornography. As deepfake technology develops, the need for technology to identify deepfake is increasing, but as one of these approaches, Facebook has developed a new technology to 'determine the characteristics of the AI model that created deepfake.' I revealed that it was inside.

Reverse engineering generative models from a single deepfake image
https://ai.facebook.com/blog/reverse-engineering-generative-model-from-a-single-deepfake-image/

Detecting the Models Behind Deepfakes --About Facebook
https://about.fb.com/news/2021/06/detecting-the-models-behind-deepfakes/

Facebook develops new method to reverse-engineer deepfakes and track their source --The Verge
https://www.theverge.com/2021/6/16/22534690/facebook-deepfake-detection-reverse-engineer-ai-model-hyperparameters

Since 2018, 'fake pornography', which was created by artificial intelligence machine learning the face of a famous actress and combining it with existing pornographic images, as if the person himself was appearing, has exploded.

AI-made 'Pornography of Famous Actresses' Explodes-GIGAZINE

Since then, the technology to generate existing human-like images and videos using machine learning tools has made remarkable progress, and there have been many cases of causing mental damage to people, which has become a problem. For this reason, some regions have introduced legislation.

There is also the problem that the reliability of information is greatly impaired by making it possible to generate high-precision fake images with artificial intelligence. In addition to fake pornography, it is also possible to make deep fake of celebrities such as Facebook CEO Mark Zuckerberg and President Trump to make statements and events that are not actually done. The more prominent the target of deepfake, the greater the impact.

For this reason, in recent years, there has been an urgent need to develop technology that can distinguish fake images and fake images. The newly announced ' How to reverse engineer deepfake' is to analyze the image generated by AI and determine the characteristics of the machine learning model that created the image.

Past studies have also made it possible to identify known machine learning models from fake images. However, according to Facebook's Tal Hassner, deepfake software can be easily customized. For this reason, deepfake generators may customize their models to erase their traces. On the other hand, the new method announced by Facebook is that even machine learning models that have not been announced so far can grasp their characteristics from the generated images.

You can see the specific mechanism from the following movie.

Reverse engineering generative models from a single deepfake


Facebook's new technology is believed to be able to detect that the creators of images uploaded to multiple locations are the same, as well as show evidence of being the culprit when the suspect's PC is seized. I will.

However, it should be noted that this technology is still in the research stage. The Verge, an overseas news media, pointed out that the reliability of these technologies is not sufficient, citing that the detection rate of the algorithm that won the deepfake detection contest in 2020 was 65.18%.