Year: 2023
Type: dissertation
Abstract: Neural networks are vulnerable to adversarial examples. In this thesis, we propose a set of defense strategies allowing a protected application. We show that the activation values of neural networks carry robustness-sensitive information. Based on this insight, we first present an architecture allowing a reliable detection of adversarial examples. We show that this architecture can further be adapted to be used in anomaly detection. Finally, we present means to leverage an analysis of the models' activations during neural-network training. This allows us to increase robustness.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7008892152