Most of the information I found while researching came from the articles “Learning Securely” by Erica Klarreich and “Model Learning” by Fritz Vaandrager. When there was something tough I didn’t immediately understand (which happened several times) the short summaries of ML topics in the DARPA videos on AI were very helpful and filled in the gaps.
Study into machine learning is advancing incredibly quickly, but nobody has really thought about protecting AI from hackers yet. In “Learning Securely” Klarreich demonstrates how naive machine learning programs are, and that hackers can trick an intelligent program into believing things that are false by looking at the way that machine studied and editing the input they feed the machine appropriately.
When children are learning to do something they build a mental model of the task in their head. Without really thinking about it, they keep track of all the ways they’ve tried to do the task before and all the ways it has ended up so far. In “Model Learning” Vaandrager thinks that machines should be taught to learn new tasks the same way. Using black-box state diagrams a machine can collect what it has learned by messing around and turn that information into something useful.
Study into machine learning is advancing incredibly quickly, but nobody has really thought about protecting AI from hackers yet. In “Learning Securely” Klarreich demonstrates how naive machine learning programs are, and that hackers can trick an intelligent program into believing things that are false by looking at the way that machine studied and editing the input they feed the machine appropriately.
When children are learning to do something they build a mental model of the task in their head. Without really thinking about it, they keep track of all the ways they’ve tried to do the task before and all the ways it has ended up so far. In “Model Learning” Vaandrager thinks that machines should be taught to learn new tasks the same way. Using black-box state diagrams a machine can collect what it has learned by messing around and turn that information into something useful.