Resisting Adversarial Examples via Wavelet Extension and Denoising

  • Qinkai Zheng
  • , Han Qiu
  • , Tianwei Zhang
  • , Gerard Memmi
  • , Meikang Qiu
  • , Jialiang Lu

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

Abstract

It is well known that Deep Neural Networks are vulnerable to adversarial examples. An adversary can inject carefully-crafted perturbations on clean input to manipulate the model output. In this paper, we propose a novel method, WED (Wavelet Extension and Denoising), to better resist adversarial examples. Specifically, WED adopts a wavelet transform to extend the input dimension with the image structures and basic elements. This can add significant difficulty for the adversary to calculate effective perturbations. WED further utilizes wavelet denoising to reduce the impact of adversarial perturbations on the model performance. Evaluations show that WED can resist 7 common adversarial attacks under both black-box and white-box scenarios. It outperforms two state-of-the-art wavelet-based approaches for both model accuracy and defense effectiveness.

Original languageEnglish
Title of host publicationSmart Computing and Communication - 5th International Conference, SmartCom 2020, Proceedings
EditorsMeikang Qiu
PublisherSpringer Science and Business Media Deutschland GmbH
Pages204-214
Number of pages11
ISBN (Print)9783030747169
DOIs
Publication statusPublished - 1 Jan 2021
Event5th International Conference on Smart Computing and Communication, SmartCom 2020 - Paris, France
Duration: 29 Dec 202031 Dec 2020

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume12608 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference5th International Conference on Smart Computing and Communication, SmartCom 2020
Country/TerritoryFrance
CityParis
Period29/12/2031/12/20

Keywords

  • Adversarial examples
  • Deep Learning
  • Image denoising
  • Model robustness
  • Wavelet transform

Fingerprint

Dive into the research topics of 'Resisting Adversarial Examples via Wavelet Extension and Denoising'. Together they form a unique fingerprint.

Cite this