Erlangen 2026 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

T: Fachverband Teilchenphysik

T 71: Data, AI, Computing, Electronics VII

T 71.2: Talk

Thursday, March 19, 2026, 16:30–16:45, KH 00.024

Utilizing Adversarial Training for IceCube's Advanced Northern Track Selection — •Marco Zimmermann, Shuyang Deng, Lasse Düser, Philipp Soldin, Sönke Schwirn, and Christopher Wiebusch — Rwth Aachen

IceCube is a neutrino observatory at the South Pole equipped with over 5000 photomultiplier tubes (PMTs), capable of detecting Cherenkov light from neutrino and cosmic-ray induced muons. The Advanced Northern Track Selection (ANTS) differentiates between these two types of muons with a deep neural network approach. The ANTS network first encodes the charge-time series from each PMT into ten abstract features via a Transformer. These features then serve as input for graph neural networks, which perform the aforementioned differentiation as well as energy and directional reconstructions, and event topology classification. To improve robustness, a network can be trained with adversarial attacks, where the input is modified by adding minimal perturbations with the aim of producing incorrect outputs. We will discuss the application of adversarial training of the ANTS' networks, with the example of the event topology classifier, and present methods to visualize the effect of the added perturbations.

Keywords: Deep Learning; IceCube; Adversarial Attacks; Neural Networks