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.