Seminario Datalab
Can Small Data Manipulations Compromise Machine Learning-Based Decisions?
Ponente: Roi Naveiro (CUNEF Universidad)Fecha: jueves 20 de marzo de 2025 - 12:00Lugar: Aula Gris 1, ICMAT
Resumen:
Adversarial machine learning research has shown that statistical models are vulnerable to malicious data manipulation. In particular, decisions based on probabilistic machine learning models can be compromised by strategically deleting or replicating just a small fraction of data points. We propose a general strategy for identifying minimal manipulations capable of steering the Bayesian posterior toward an adversarially chosen target distribution—one that leads to incorrect decisions. Crucially, our approach remains effective even when a closed-form posterior is unavailable and only sampling access is provided.
