SMuK 2023 – wissenschaftliches Programm

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

T: Fachverband Teilchenphysik

T 10: ML Methods I

T 10.1: Vortrag

Montag, 20. März 2023, 16:30–16:45, HSZ/0405

Fooling IceCube's Deep Neural Networks — •Oliver Janik, Markus Bachlechner, Thilo Birkenfeld, Philipp Soldin, Christopher Wiebusch, and Katharina Winkler for the IceCube collaboration — III. Physikalisches Institut B, RWTH Aachen University

Deep neural networks (DNNs) find more and more use in the data analysis of physics experiments. In IceCube, such networks are used as classifiers for particle identification or as regressors to reconstruct the direction and energy of particles. In the context of adverserial attacks, it has been observed that imperceptible changes to the input of DNNs can alter the output drastically. Algorithms like DeepFool can calculate minimal changes of the input in order to obtain a wrong output, thus fooling the network. This talk will focus on testing the robustness of IceCube's DNNs to such minimal changes.