Göttingen 2025 – wissenschaftliches Programm

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

AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz

AKPIK 3: Machine Learning in Particle- and Astroparticle Physics

AKPIK 3.4: Vortrag

Donnerstag, 3. April 2025, 17:00–17:15, Theo 0.134

Advanced Northern Tracks Selection using a Graph Convolutional Neural Network for the IceCube Neutrino Observatory: Adversarial Training — •Leon Hamacher, Philipp Behrens, Jakob Böttcher, Shuyang Deng, Lasse Düser, Philipp Fürst, Philipp Soldin, and Christopher Wiebusch for the IceCube collaboration — RWTH Aachen

The IceCube Neutrino Observatory, located at the South Pole, detects atmospheric and astrophysical neutrinos. One important task is differentiating between these neutrinos and muons induced by air showers. The Advanced Northern Tracks Selection (ANTS) accomplishes this identification using a graph-convolutional neural network. However, neural networks can be sensitive to minor adversarial perturbations, which can significantly alter their outputs. Adversarial training is a method to include artificially perturbed data during the training process to enhance resistance to such perturbations. For this purpose, a dedicated algorithm, MiniFool, has been developed that takes experimental uncertainties into account. This talk presents the results of applying MiniFool to ANTS.

Keywords: IceCube; Advanced Northern Tracks Selection; Deep Learning; Adversarial Training