Sure a deepfake is convincing on a low resolution compressed video, but what about a 4k or 8k video where lens distortion and fine details are everywhere? I don't have a lot of confidence that facial pores could be convincingly simulated.
https://copyrightimage.com/2018/05/09/better-image-upscaling...
This is from 2018, which is an eternity for machine learning research.
Here's a random example:
https://ieeexplore.ieee.org/document/6619284
They can detect a persons pulse from tiny variations in head movement, I'd like to see a deepfake simulate that! And it's not about this one thing, I'm sure you could design against these tiny measurements one by one... the key is that there are a lot of them and a deepfake is going to have to simulate an enormous amount of data in order to pass what will become trivial tests for realism. This is made a much easier weapon against deepfakes when you have progressively better videos.
Let's say there was video recording of a politician in a hotel in Russia, to pick a random example. We would already expect it to be low-quality video even if it was real.
But why? The cheap last-gen iPhone in my pocket can take 4k videos, almost everybody has something like it in their pockets. Quality is cheap and will continue spreading, fuzzy video that looks like it was taken on a 90s camcorder will itself seem suspicious.