SMuK 2023 – wissenschaftliches Programm

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

T: Fachverband Teilchenphysik

T 34: ML Methods II

T 34.3: Vortrag

Dienstag, 21. März 2023, 17:30–17:45, HSZ/0405

Improving the robustness of jet tagging algorithms with adversarial training — •Hendrik Schönen¹, Annika Stein¹, Judith Bennertz¹, Xavier Coubez^1,2, Alexander Jung¹, Summer Kassem¹, Ming-Yan Lee¹, Spandan Mondal¹, Alexandre de Moor³, Andrzej Novak¹, and Alexander Schmidt¹ — ¹III. Physikalisches Institut A, RWTH Aachen University, Germany — ²Brown University, USA — ³Vrije Universiteit Brussel, Belgium

Neural network architectures have advanced over the last decade and are an important part of current jet flavour tagging algorithms. Since these algorithms rely on training the network with simulated events as input, they might have a worse performance on detector data due to data/MC deviations. A possible approach to address this issue is adversarial training, which uses distorted inputs for training. One possibility to distort the inputs is applying a FGSM attack, which shifts the inputs in a way that maximizes the loss with a fixed magnitude. This talk is about the impact of adversarial training on the model performance and robustness.