ABBA Neural Networks: Coping with Positivity, Expressivity, and Robustness - Inria - Institut national de recherche en sciences et technologies du numérique Access content directly
Journal Articles SIAM Journal on Mathematics of Data Science Year : 2024

ABBA Neural Networks: Coping with Positivity, Expressivity, and Robustness

Abstract

We introduce ABBA networks, a novel class of (almost) non-negative neural networks, which are shown to possess a series of appealing properties. In particular, we demonstrate that these networks are universal approximators while enjoying the advantages of non-negative weighted networks. We derive tight Lipschitz bounds both in the fully connected and convolutional cases. We propose a strategy for designing ABBA nets that are robust against adversarial attacks, by finely controlling the Lipschitz constant of the network during the training phase. We show that our method outperforms other state-of-the-art defenses against adversarial white-box attackers. Experiments are performed on image classification tasks on four benchmark datasets.
Fichier principal
Vignette du fichier
siamonline_220329.pdf (7.03 Mo) Télécharger le fichier
supplemental.pdf (5.59 Mo) Télécharger le fichier
Origin : Files produced by the author(s)
Origin : Files produced by the author(s)

Dates and versions

hal-04386260 , version 1 (10-01-2024)

Licence

Attribution - NonCommercial

Identifiers

  • HAL Id : hal-04386260 , version 1

Cite

Ana-Antonia Neacșu, Jean-Christophe Pesquet, Vlad Vasilescu, Corneliu Burileanu. ABBA Neural Networks: Coping with Positivity, Expressivity, and Robustness. SIAM Journal on Mathematics of Data Science, In press. ⟨hal-04386260⟩
101 View
102 Download

Share

Gmail Facebook X LinkedIn More