Year: 2025
Type: dissertation
Abstract: This thesis addresses adversarial robustness of neural networks in three different contexts. First, we propose an approach to verify (non-)robustness, based on the approximated reachable set of a network. Second, we propose a new attack type to analyze robustness of uncertainty models that provide estimates indicating how certain they are about a prediction. Third, we analyze how adversarial robustness behaves during transfer learning by analyzing predictions and latent representations.
Authors: Kopetzki, Anna-Kathrin, Günnemann, Stephan 1983-, Technische Universität München
Institution: Technical University of Munich
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: Strong
OpenAlex ID: W6979761941
