That actually leads to a very interesting point; would self driving cars be vulnerable to adversarial imagery?
Neural nets are well known for being easily fooled [1] ... I wonder if you could create similar situations for self-driving cars.
[1] http://www.evolvingai.org/fooling