Communication Dans Un Congrès Année : 2024

Neuron Pairs in Binarized Neural Networks Robustness Verification via Integer Linear Programming

Résumé

In the context of classication, robustness verication of a neural network is the problem which consists in determining if small changes of inputs lead to a change of their assigned classes. We investigate such a problem on binarized neural networks via an integer linear programming perspective. We namely present a constraint generation framework based on disjunctive programming and complete descriptions of polytopes related to outputs of neuron pairs. We also introduce an alternative relying on specic families of facet dening inequalities. Preliminary experiments assess the performance of the latter approach against recent single neuron convexication results.
Fichier principal
Vignette du fichier
DLJN5.pdf (444.61 Ko) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04890350 , version 1 (23-01-2025)

Identifiants

Citer

Dymitr Lubczyk, José Neto. Neuron Pairs in Binarized Neural Networks Robustness Verification via Integer Linear Programming. International Symposium on Combinatorial Optimization, May 2024, Tenerife (Canary Islands), Spain. pp.305-317, ⟨10.1007/978-3-031-60924-4_23⟩. ⟨hal-04890350⟩
0 Consultations
0 Téléchargements

Altmetric

Partager

More