SEMINARIO DE ESTADÍSTICA

Lunes 25 de noviembre de 2024, 12:00 horas

David Ríos Insua (CSIC-ICMAT)

Sala 520, Departamento de Matemáticas (UAM)

TÍTULO: Adversarial Machine Learning

ABSTRACT.- Recent advances in Artificial Intelligence and Machine Learning (ML) have revolutionized the use of automated systems for making predictions and supporting decisions in complex areas like drug design, automated driving, and medical decision-making, to name but a few. These technologies offer numerous benefits, but they also present challenges, particularly in terms of security:  adversaries may attempt to fool these systems with malicious purposes to attain a benefit. This has given rise to the emerging field of Adversarial Machine Learning (AML) which focuses on three topics: 1) designing attacks against AML systems; 2) designing defenses against such attacks; 3) global frameworks/pipelines to robustify ML algos. Most current advancements in AML aim to protect ML algorithms from worst-case attacks with a characteristic game theoretic flavor. However, these attacks are often unrealistic because they assume a level of common knowledge between the ML system designers and attackers, which is often unrealistic. I shall provide an overview on key concepts and approaches in AML and then outline some improvements over current methods with a Bayesian perspective.