Deepfake is an amalgamation of the words ‘deep learning’ and ‘Fake’. In videos it is where the face of another person is stitched onto the original person in that video. The technology uses a type of artificial intelligence which learns the idiosyncrasies and gestures of a particular face and is therefore able to convincingly recreate them to do any movement.
Audio can also be deepfaked, this is often used to scam people into thinking someone has called them when really the voice is a fake. They can also be used to make public figures sound like they have said things they never have, like the widely shared video of Barack Obama Calling Donald Trump a ‘Dipshit’.
When did it start?
Deepfakes were first really discovered in 2017 when a reddit user under the name ‘Deepfake’ posted edited porn videos onto the site where they user had swapped the faces on the porn performers for celebrities - Taylor Swift, Gal Gadot and Scarllet Johanson.
How does it work? (In simple terms)
Firstly you need a lot of images of a persons face from many different angles, you then put the images of the two peoples faces through an AI algorithm called an encoder.
It learns the similarities between the two peoples faces and compresses the images.
Then the images are then put through a decoder which is taught to recover the faces from the compressed images. Because the faces are different, you make one decoder recover the first person’s face, and another decoder to recover the second person’s face. To perform the face swap, you put the encoded images into the “wrong” decoder.
How to spot a deepfake
When the technology first started being used it was fairly easy to tell a deepfake apart from a real video. As the technology has progressed it has become much harder as people who create the videos have learnt to fix the initial issues with it.
The original videos wouldn’t blink well as understandably the images of peoples faces that were put through the algorithm would have their eyes open in the pictures. Once creators realised this they found a way to fix it.
When trying to spot a deepfake lookout for bad lip-syncing or flickering when the face turns a certain way, there should be something that gives away the fact the video has been doctored however it has become much harder to spot as the technology has advanced there are now even some apps that can be downloaded onto mobile phones that will create simple deep fakes.