…so, just feed this data back into the deep fake engine model until it is indistinguishable again?

Definitely an interesting approach to this current arms race.

something i’ve wondered is why do we need to rely on such detailed methods? surely at a pixel distribution level there’s enough tells? why is that not the case?

Very necessary, but I have no illusions that this would be a definitive solution. Just as technologies are advanced to detect these fakes, so will those who create them. Deepfake software is FOSS and therefore subject to improvements, especially if it is used by certain groups of government Devs with a lot of resources.