If you can do a statistical sample of a very small number of users, say 0.1%, and do very expensive detailed investigation of them that determines 5% are fake, you can easily extrapolate that 5% to the entire userbase with small confidence bounds, but have no idea which of the rest of the users are fake.
