I think the argument is that if you produce your fraud from an appropriate probability distribution, any "detection" method other than independently verifying the results is snake oil.