Fabrizio Costa
Exact Characterization of Safe and Adversarial Regions in ReLU Neural Networks.
Rel. Fabio Guido Mario Salassa, Marco Ghirardi. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Informatica (Computer Engineering), 2026
|
|
PDF (Tesi_di_laurea)
- Tesi
Accesso limitato a: Solo utenti staff fino al 24 Luglio 2027 (data di embargo). Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (4MB) |
Abstract
The rising adoption of deep neural networks in safety-critical domains has intensified the need for rigorous verification techniques capable of providing formal guarantees on model behavior. Within this context, local robustness is assessed by determining whether small perturbations around a correctly classified input can alter the network's prediction. If the prediction remains stable within the perturbation domain, the network is certified as safe; any input leading to a different classification constitutes an adversarial vulnerability. While most traditional verifiers provide only a binary answer (proving or disproving the existence of such violations), comparatively little attention has been devoted to the exact geometric characterization of the safe and adversarial sub-regions within the input domain.
This thesis introduces PRISM (PRuned Inverse Search Method), a novel formal verification framework for ReLU-based neural networks designed to compute the exact pre-image of output specifications
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
URI
![]() |
Modifica (riservato agli operatori) |
