I don't think you need facial recognition at all. You need a robust video hash. Need a way to map videos to a perceptual space that's independent of the resolution, the exact length and the colors. Seems very doable to me. IMO the answer here is probably to use deep learning, which would likely turn out to be very effective.

Facial recognition would be really tough to get right, but maybe the opposite would work. All faces are blurred unless uploaded by a known performer or verified account.