Heidelberg 2022 – wissenschaftliches Programm
T 26.1: Vortrag
Montag, 21. März 2022, 16:15–16:30, T-H39
Investigation of robustness of b-Tagging algorithms for the CMS Experiment — Xavier Coubez1,2, Nikolas Frediani1, Spandan Mondal1, Andrzej Novak1, Alexander Schmidt1, and •Annika Stein1 — 1RWTH Aachen University, Germany — 2Brown University, USA
Deep learning as one form of machine learning is utilized for various applications and shows its benefits also in the field of high-energy physics, or more specifically, for jet flavour tagging. However, subtle mismodelings in the simulation could be invisible to typical validation methods. Investigating the response to mismodeled input data is motivated by the later usage of the outputs in physics analyses, as the values for simulation and data are deviating. The vulnerability of b-tagging algorithms used at the CMS experiment is probed through application of adversarial attacks. In this talk, a corresponding defense strategy that improves the robustness, namely adversarial training, will be presented. Comparisons of the model performance and the susceptibility show that this method constitutes a promising candidate to reduce the vulnerability and that this could improve the capability to generalize to data.