Samsung developing algorithm needs one picture to make a fake video

The AI that can make Mona Lisa smile: Samsung reveals algorithm that creates fake talking-head videos using just ONE photo

  • A new algorithm can create faked videos using only a single picture
  • The technology uses ‘landmark’ features to bring images to life
  • In a video the algorithm animates Salvador Dali, the Mona Lisa, and more
  • Rapid advancement in the technology has caused among some

As if the world of deep-faked pictures and video wasn’t scary enough, researchers from Samsung’s AI center in Moscow have demonstrated an algorithm that can fabricate videos using only one image. 

In a video demonstration and a paper published in the pre-print journal ArXiv, the researchers show the capabilities of what is described as ‘one-shot’ and ‘few-shot’ machine learning.

The results of their system bring to life popular faces like those of surrealist painter Salvador Dali and actress Marilyn Monroe using a single still image. 

Scroll down for video 

As if the world of deep-faked pictures and video wasn’t scary enough, researchers from Samsung’s AI center in Moscow have demonstrated an algorithm that can fabricate videos using only one image. The team tested it on the Mona Lisa, for example, shown above

The more images that are fed into the program, the more realistic the resulting video becomes. 

Though a single image translated into a moving face may look noticeably altered, a sample of 32 images produces a moving picture with near lifelike accuracy. 

Researchers’ technology appears to take algorithms used to produce deepfakes far beyond existing methods not just in terms of realism, but in the amount of work that goes into creating photorealistic fakes. 

Current AI systems are usually trained on an extensive catalog, requiring the algorithm to scan large data sets of a subjects body or face. 

This method limits the system’s application to individuals who have sufficient amounts of pictures and videos available for use.

In the case of the researchers new software, however, those data sets are a more of a luxury than a necessity. 

The newest algorithm was trained using the publicly available VoxCeleb database which contains more than 7,000 images of celebrities from YouTube videos.

Using that data set, the AI was able to establish what researchers call ‘landmark’ features– universally identifiable traits — among subjects’ noise, eyes, and more.

Because of the algorithm’s recognizes common characteristics as opposed to specific traits of a subject, it’s able to quickly extrapolate images with little input and in some cases, bring to life still images of people like Fyodor Dostoevsky, who died in 1881 and was conceivably never captured on video.

This method also also means that the technology is applicable toward non-celebrities. 

Using just a few selfies from a subject, a demonstration shows how the algorithm can create a video facsimile mimicking how they might move their head and speak.

Using the data set, the AI was able to establish what researchers call ‘landmark’ features– universally identifiable traits — among subjects’ noise, eyes, and more. Its take on Salvador Dali is shown

The AI is currently only able to produce ‘talking head’ style videos of subjects from the shoulders up, but if the algorithm is anything like similar deepfake systems, that could quickly change

In a video demonstration and a paper published in the pre-print journal ArXiv, the researchers show the capabilities of what is described as ‘one-shot’ and ‘few-shot’ machine learning

The AI is currently only able to produce ‘talking head’ style videos of subjects from the shoulders up, but if the algorithm is anything like similar deepfake systems, that could quickly change.

Another recent and startling technology recently demonstrated by former employees of Google and Microsoft shows how an algorithm is able to convincingly replicate the voice of popular podcaster, Joe Rogan.

Skeptics of deepfake technology say those two technologies combined would allow bad actors to feasibly fake videos that could spread misinformation on politicians or steal other victims’ identity. 

Deepfaking technology has already been used at length to superimpose the face of celebrities on the bodies of porn actors causing some to call for legislation that would ban the use.

WHY IS FAKE CELEBRITY PORN MADE BY AN AI SO CONCERNING?

Back in December, it was discovered that Reddit users were creating fake pornography using celebrity faces pasted on to adult film actresses’ bodies. 

The disturbing videos, created by Reddit user deepfakes, look strikingly real as a result of a sophisticated machine learning algorithm, which uses photographs to create human masks that are then overlaid on top of adult film footage. 

Now, AI-assisted porn is spreading all over Reddit, thanks to an easy-to-use app that can be downloaded directly to your desktop computer, according to Motherboard. 

The video, created by Reddit user deepfakes, features a woman who takes on the rough likeness of Gadot, with the actor’s face overlaid on another person’s head. A clip from the video is shown

To create the likeness of Gal Gadot, for example, the algorithm was trained on real porn videos and images of actor, allowing it to create an approximation of the actor’s face that can be applied to the moving figure in the video.

As all of this is freely available information, it could be done without that person’s consent. 

And, as Motherboard notes, people today are constantly uploading photos of themselves to various social media platforms, meaning someone could use such a technique to harass someone they know.

Source: Read Full Article