VERY interesting questions! Unfortunatly I can't answer the questions, since I don't know enough about them.

To make the literature search easier: Your second cocern is called "poisioning attacks" and is one of the problems "adversarial machine learning" is concerned with.