Security in machine learning: a cryptographer's perspective

Speaker: Giorgia Marson, NEC Labs Europe, Germany

Abstract:
Recent advances in AI have made machine-learning technologies ubiquitous in many applications and services. However, current ML algorithms are extremely fragile to adversarial manipulations, which makes them unreliable (and therefore unsuitable) in adversarial settings. The focus of this talk will be on vulnerabilities of ML classifiers to evasion and poisoning attacks. I will present existing approaches to defend against these attacks, highlighting the challenges in securing ML classifiers, and discuss open problems and possible future directions from my own perspective as a cryptographer.

Biography:
Giorgia Marson is a senior researcher in the Security Group at NEC Labs Europe, Germany. She received her M.Sc. in Mathematics from Sapienza University of Rome, Italy, and her PhD from TU Darmstadt, Germany. During her PhD time, she contributed research on provably-secure cryptography with focus on modeling and analyzing the security of cryptographic channels. Afterwards she has worked as postdoc at Ruhr University Bochum, Germany, as research scientist at NEC Labs Europe, and as postdoc at University of Bern, Switzerland, broadening her research interests towards distributed systems and system security.

Host: Prof. Patrick Eugster