| Meno: | Paulína |
|---|---|
| Priezvisko: | Jaremčuková |
| Názov: | Complexity of trained neural networks' decision boundaries, and its relationship with the adversarial robustness |
| Vedúci: | Mgr. Iveta Bečková, PhD. |
| Rok: | 2026 |
| Kµúčové slová: | neural network, adversarial example, decision boundary, boundary instance |
| Abstrakt: | Modern neural networks remain vulnerable to small, deliberate perturbations known as adversarial examples. Since their discovery in 2014, several hypotheses about the underlying cause have been proposed; however, no clear consensus has been reached yet. This work analyses robustness of neural networks in relation to the complexity of their decision boundaries. We achieve this by evaluating trained models using methods from prior research, measuring local curvature and boundary irregularity, and assessing their robustness against dierent adversarial attacks. In addition, we examine how adversarial examples dier from randomly sampled boundary instances and the clean images. The aim of this work is to provide the insight into the behavior of decision boundary in dierent parts of input space, linking its geometric properties with adversarial robustness. |
Súbory diplomovej práce:
| codes.zip |
| main.pdf |
Súbory prezentácie na obhajobe: