The argument in favor of these schemes is that tasks like object recognition are usually much more complicated to carry out than text recognition and thus should be far more resilient to machine studying centered attacks.