I have a similar background (worked about a year on some anti-abuse systems), and completely agree.

This is an incredibly hard problem space. Machine learning is already hard before you start introducing adversaries...