Small Deepfake Artefacts

Final Work:

In my final work, I experimented with the technology of deepfakes.
With AI, or more precisely with a neuronal network it is possible to fake real footage so that it looks like someone else has made e.g. a speech. You teach a neural network to recognize human faces, for example, by sending hundreds or thousands of photographs and providing each image with the information whether it may or may not be a face. The network recognizes parts that make up a human face. The network, as soon as it sees a new picture, can calculate to what extent the characteristics are present, and then make the decision „This is a face“ - or not.

If you provide the network with enough data, it‘s hard to tell afterwards if the footage is real or not.

I wanted to create some deepfakes to understand the process and the workflow to create a faked video.
In the video above you can see my process to produce the final outcome. I would say I'm still in the learning process because its a pretty complex and time-consuming topic. I decided pretty early to use the LIAEF128 training model because the fast results looked better than other models but I still want to give the others a chance and would like to see how the results are going to be with different models. But so far I have a basic understanding of how deepfakes are working.

Deepfake | John F. Kennedy - Civil Rights Speech
16769 Iterations
LIAEF128 Training Model
Batchsize: 8

Deepfake | Elon Musk - Ted Talk
29100 Iterations
LIAEF128 Training Model
Batchsize: 16