Göttingen 2025 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

T: Fachverband Teilchenphysik

T 33: Data, AI, Computing, Electronics III (ML in Jet Tagging, Misc.)

T 33.5: Vortrag

Dienstag, 1. April 2025, 17:15–17:30, VG 2.101

Adversarial Studies on Jet-Flavor Tagging Machine Learning Algorithms using PAIReD Jets within the CMS Experiment — Alexander Jung¹, Spandan Mondal², Alexander Schmidt¹, Jan Schulz¹, and •Ulrich Willemsen¹ — ¹III. Physikalisches Institut A, RWTH Aachen — ²Brown University

The PAIReD tagger is a novel jet flavor tagging algorithm in CMS that employs unconventional large-radius jets to identify Higgs boson decays to pairs of heavy-flavor quarks. In this talk, the vulnerability of machine learning-based jet flavor taggers to adversarial attacks is investigated, with a focus on the ParticleTransformer architecture used in the PAIReD tagger. It is shown that this architecture is more susceptible to adversarial perturbations than other established models. To mitigate this vulnerability, adversarial training is applied, incorporating adversarial examples into the training process. It is demonstrated that adversarial training enhances the robustness of the PAIReD tagger, recovering almost the nominal performance on both undisturbed and attacked inputs. These findings provide valuable insights into the behavior of the PAIReD tagger and the ParticleTransformer architecture for future applications in the CMS experiment.

Keywords: Machine Learning; Adversarial Attacks; ParticleTransformer; Jet Flavor Tagging; PAIReD tagger