Skip to main navigation Skip to search Skip to main content

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

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

In the context of classification, robustness verification 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 specific families of facet defining inequalities. Preliminary experiments assess the performance of the latter approach against recent single neuron convexification results.

Original languageEnglish
Title of host publicationCombinatorial Optimization - 8th International Symposium, ISCO 2024, Revised Selected Papers
EditorsAmitabh Basu, Ali Ridha Mahjoub, Ali Ridha Mahjoub, Juan José Salazar González
PublisherSpringer Science and Business Media Deutschland GmbH
Pages305-317
Number of pages13
ISBN (Print)9783031609237
DOIs
Publication statusPublished - 1 Jan 2024
Event8th International Symposium on Combinatorial Optimization, ISCO 2024 - La Laguna, Spain
Duration: 22 May 202424 May 2024

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume14594 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference8th International Symposium on Combinatorial Optimization, ISCO 2024
Country/TerritorySpain
CityLa Laguna
Period22/05/2424/05/24

Keywords

  • Cutting-plane
  • Disjunctive programming
  • Robustness verification

Fingerprint

Dive into the research topics of 'Neuron Pairs in Binarized Neural Networks Robustness Verification via Integer Linear Programming'. Together they form a unique fingerprint.

Cite this