Watch: AI-created Elon Musk Zoom-bombs virtual meeting
What's the story
You may already know about deepfakes, the AI-generated clips that show one person's face or voice in another person's video. The technology behind such deceptive videos has grown insanely popular, but just recently, a researcher showcased something different - real-time deepfake creation. More interestingly, they used this technology to make a "fake Elon Musk" Zoom-bomb an ongoing virtual meeting. Here's more about it.
Surprise
Musk appears in virtual meeting
Using his new deepfake project, programmer Ali Aliev jumped as Musk in a three-way Zoom meeting. The tool superimposed the face of Musk over Aliev's in real-time, which, thanks to the frame drops and lower resolution of the Zoom call, made him look just like the tech mogul. He looked so real that the other two participants of the call were caught off guard.
Work
So, how Aliev did it?
Aliev transformed himself into Musk by creating Avatarify, a program that uses an AI algorithm to place the face of any person over yours - in real-time. It does not require any pre-recorded segments and continues to insert the target's face over the source even when the clip streams online, which is not common in the case of deepfakes.
Details
Use of "First Order Motion Model For Image Animation"
Aliev says he developed Avatarify by using the open-source code from the "First Order Motion Model For Image Animation" - a face-swapping program developed jointly by the University of Trento in Italy and the team at Snap. Notably, the tool doesn't require any training and can start working as soon as you place the target images in the 'avatars' folder of the program directory.
Requirement
You would need some hefty hardware to use it
Once you set up the program and add the images, you can pretend to be anyone - from Albert Einstein to Steve Jobs to someone else - on Zoom, Skype, or Slack. However, do note that the tool requires gaming PC-level hardware and will not work smoothly on a mediocre machine. All the elements, including the software code and instructions, are available on Github.