Out of curiosity- why is detection of fake videos a hard problem? It sounds like a classification problem that a simple NN with enough data would be able to solve(?)
A lot of the scariest work in fake video lately has involved generative adversarial networks (GANs) whose whole optimization strategy as I understand it is to try to produce videos that can fool classifier/discriminator NNs, which are trained in parallel.
Some approaches target multiple models to come up with more general attacks. It's also worth noting that manipulations that will fool humans are starting to pop up as well.