Theorie, Einschränkungen und Anwendungen von Verfahren zur Robustheitsverifikation neuronaler Netzwerke

mediaTUM – the media and publications repository of the Technical University Munich (Technical University Munich) 2026 Thesis Open access

Read full text
Authors: Aleksei Kuvshinov
Cited by: 0

Abstract

This dissertation addresses key challenges in ensuring the reliability of machine learning systems, focusing on adversarial robustness verification. We propose novel methods for robustness certification via convex relaxation, analyze the limitations of randomized smoothing in high dimensions, and explore data dependent smoothing techniques. Finally, we apply generative models to real-world domain reconstruction for car key localization, bridging theory and application.

Fields of Study & Subjects

Adversarial Robustness in Machine Learning Formal Methods in Verification Software Testing and Debugging Techniques

Publication details

Journal / source
mediaTUM – the media and publications repository of the Technical University Munich (Technical University Munich)
Publication date
2 February 2026
Language
EN
Document type
Thesis