Abstract

The widespread adoption of AI, and more specifically of ML, has profoundly impacted society. Few would dispute this, especially given the rise of Large Language Models (LLMs). Beyond these, cutting edge applications of ML span autonomous driving, drug discovery, or predictive medical systems, to name but a few. Yet the full risks of deploying these tools at scale remain poorly charted. Our focus in the talk refers to one of those risks, the security of ML systems, which has led to the relatively recent field of adversarial machine learning (AML). However, the prevailing paradigm within AML has been game-theoretic affecting its three main problems of interest: designing attacks to ML systems, deriving defenses protecting ML systems, and producing pipelines to robustify such systems against adversarial manipulations. In the talk I shall discuss the relevance of Bayesian ideas in the field of AML and discuss recent developments.