Black Boxes and Black Hats

Casimir Kulikowski (Computer Science)

Artificial Intelligence (AI), and its most popular recent offspring “Deep Learning” hold many promises for improving our lives. “Internet of Things” (IoT) devices could use such techniques in new ways that we can barely imagine. But “deep learning” systems are generally “black boxes” whose internal workings are something of a mystery, even to their designers. We often don’t exactly know how they do what they do. Interactions between ‘black box’ devices (and exploitation of security vulnerabilities in them) is a problem that has barely begun to be studied. How can we repair a system that we cannot fully describe that has been damaged, or held for ransom by the bad guys? This seminar will review the challenges that arise for security and safety of inter-connected systems of “smart, but black” boxes that make up the IoT, and discuss how AI systems will have to be changed in order to make them more transparent in their performance and easily and independently tested and verified.