The argument in favor of these techniques is that tasks like object recognition are generally more complicated to carry out than text recognition and thus should be far more resilient to equipment learning dependent attacks.