[yczsdnxs] Neural net backdoors

A vendor provides you with a neural network system that identifies bad entities, for some definition of "bad".  Does it contain a backdoor?  It is currently very difficult to look at a neural network and determine what is does.  How would one design a difficult-to-detect backdoor into a deep neural net system?  How would one prove a system doesn't have a backdoor?

Of course, the ultimate example of such a "system" is a double-agent human whose brain is the neural network.