If you've trained one you've trained them all: inter-architecture similarity increases with robustness.