Fraunhofer Institute for Industrial Mathematics ITWMKaiserslautern, Fraunhofer ITWM / Lecture / Machine and Deep Learning Seminar / July 06, 2023, 11:00 – 12:30 p.m.
Breaking the Boundaries of Certified Robustness
Speaker: Dr. Andrew Cullen (Department of Adversarial Machine Learning, University of Melbourne)
Fraunhofer ITWM Fellow
Science thrives on the exchange of ideas, not only with colleagues on site. Impulses from outside are just as important. This is why the Fraunhofer Institute for Industrial Mathematics ITWM has established a Fellows Programme: Prominent guest scientists come to the institute for a maximum of six months and work on a topic together with our researchers. Lorenz T. Biegler is one of our Fraunhofer ITWM Fellows.
Abstract: Breaking the Boundaries of Certified Robustness
Defences against adversarial attacks are typically crafted responses to a particular type of attack. However, by their very nature such defences can be defeated through the creation of different attack vectors. In contrast to this attacker-defender paradigm, certification mechanisms produce guarantees of robustness against all possible adversarial attacks, up to a calculated norm-bounded size. In this talk I'll introduce certification mechanisms for Machine Learning, before introducing new works by my team that have demonstrated both how certified guarantees can be improved upon, and how they can be exploited.