One of the aspects not discussed yet in our analysis is the impact of calculation time to the quality of a Deepfake. All examples discussed so far took 24 hours to generate. All discussions in this article are based on the material used in the case of Source 5.000 images ⇒ Destination 5.000 images. It must be noted that the pre-trained model is based on Donald Trump and Nicolas Cage.

Impact of Calculation Time

Abort after 1 second

The result shows clearly that Nicolas Cage is used as source model. The face appears a bit unsharp and blurry.

Abort after 1 minute

The rendering of 1 minute shows a first signs of changes. In comparison to the first video cancelled after just one second this one indicates a slight change in face color and a tiny bit of restructuring of the facial shape, especially in regards of the mouth.

Abort after 1 hour

There are more improvements, the original face of Nicolas Cage is not that obvious anymore. But there is still no George Clooney visible yet.

Abort after 6 hours

After six hours we are able to dermine obvious changes. The generated face is more round, less long and not that much looking like Nicolas Cage anymore. There are also changes in regard of the eyes. They are much bigger than in the earlier calculations.

Abort after 12 hours

It is not possible to see changes beetween the rendering time of six and twelve hours.

Abort after 24 hours

There are even more differences between the videos calculating six hours and 24 hours. The 24 hour long rendering made the eyes a bit bigger and the lips a bit fuller.

Abort after 48 hours

There are no visible changes beetween the calculation time of 24 hours and 48 hours.

Abort after 1 week

There is even no visible changes between 24 hours and one week calculation time.

Future of Deepfakes

Since we have started our analysis there were serious improvements in the field of Deep Learning. In a new paper with the title Few-Shot Adversarial Learning of Realistic Neural Talking Head Models it was explained how to model a realistic head with just eight images.

Conclusion

Our analysis made it clear that a pre-trained model might just require 24 hours to create a solid result. More training than these 24 hours does not really improve the results significantly. In regards to all our tests we may conclude that it is suggested to use 500 images, include all angles and respect the lighting to train for 24 hours to generate a good Deepfake.

About the Author

Andrea Hauser graduated with a Bachelor of Science FHO in information technology at the University of Applied Sciences Rapperswil. She is focusing her offensive work on web application security testing and the realization of social engineering campaigns. Her research focus is creating and analyzing deepfakes. (ORCID 0000-0002-5161-8658)

Links

• https://arxiv.org/pdf/1905.08233v1.pdf