Lecture: The Insecurity of Deep Neural Networks

Adi Shamir

Machine learning has made tremendous progress in the last decade, solving a broad range of tasks ranging from recognizing objects to chatting in natural language. However, today's amazing systems have a dark side: They are incredibly easy to fool by malicious actors. In this talk, I will describe some of these attacks, such as adversarial examples and trapdoored systems, and try to solve the mystery of what makes systems which are so powerful so vulnerable.