Escaping Adversarial Attacks with Egyptian Mirrors

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

Abstract

Adversarial robustness received significant attention over the past years, due to its critical practical role. Complementary to the existing literature on adversarial training, we explore weight-space ensembles of independently trained models. We propose a defense against adversarial examples which takes advantage of the latest empirical findings on linear mode connectivity of overparameterized models modulo permutation invariance. Egyptian Mirrors defense escapes adversarial attacks by moving along linear paths between pair-wise aligned functionally diverse models, while frequently and arbitrary changing ensembling direction. We evaluate the proposed defense using adversarial examples generated by FGSM and PGD attacks and show improvements up to 8% and 33% test accuracy on 2-layer MLP and VGG11 architectures trained on GTSRB and CIFAR10 datasets respectively.

Original languageEnglish
Title of host publicationACM MobiCom 2023 - Proceedings of the 2023 2nd ACM Workshop on Data Privacy and Federated Learning Technologies for Mobile Edge Network, FedEdge 2023
PublisherAssociation for Computing Machinery (ACM)
Pages131-136
Number of pages6
ISBN (Electronic)9798400703447
DOIs
Publication statusPublished - 12 Mar 2024
Event2nd ACM Workshop on Data Privacy and Federated Learning Technologies for Mobile Edge Network, FedEdge 2023, Collocated with ACM MobiCom 2023 - Madrid, Spain
Duration: 2 Oct 20232 Oct 2023

Publication series

NameACM MobiCom 2023 - Proceedings of the 2023 2nd ACM Workshop on Data Privacy and Federated Learning Technologies for Mobile Edge Network, FedEdge 2023

Conference

Conference2nd ACM Workshop on Data Privacy and Federated Learning Technologies for Mobile Edge Network, FedEdge 2023, Collocated with ACM MobiCom 2023
Country/TerritorySpain
CityMadrid
Period2/10/232/10/23

ASJC Scopus subject areas

  • Artificial Intelligence
  • Computer Networks and Communications
  • Safety, Risk, Reliability and Quality

Fields of Expertise

  • Information, Communication & Computing

Fingerprint

Dive into the research topics of 'Escaping Adversarial Attacks with Egyptian Mirrors'. Together they form a unique fingerprint.

Cite this