I cannot disagree with you but the fact is and remains that the field is a budding one and this is but a building block towards better understanding the nature of using sets of geometric transformations for pattern recognition in A.I.
If this was to be directly implemented in the real world your arguments would need to be addressed, however I'm certain this research paper wouldn't be abused in that manner.
Disclaimer: I'm a graduate student studying Computer Vision.
Google trained their AI Go algorithm by playing against itself. It kind of has the same feel to it.
Edit: oops, turns out I mistakenly responded to the content of the paper instead of the fact that it exists and the form of its existence. Sorry.