Towards robust neural networks via random self-ensemble

Xuanqing Liu, Minhao Cheng, Huan Zhang, Cho Jui Hsieh

Research output: Chapter in Book/Report/Conference proceedingConference contribution

40 Scopus citations

Abstract

Recent studies have revealed the vulnerability of deep neural networks: A small adversarial perturbation that is imperceptible to human can easily make a well-trained deep neural network misclassify. This makes it unsafe to apply neural networks in security-critical applications. In this paper, we propose a new defense algorithm called Random Self-Ensemble (RSE) by combining two important concepts: randomness and ensemble. To protect a targeted model, RSE adds random noise layers to the neural network to prevent the strong gradient-based attacks, and ensembles the prediction over random noises to stabilize the performance. We show that our algorithm is equivalent to ensemble an infinite number of noisy models fϵ without any additional memory overhead, and the proposed training procedure based on noisy stochastic gradient descent can ensure the ensemble model has a good predictive capability. Our algorithm significantly outperforms previous defense techniques on real data sets. For instance, on CIFAR-10 with VGG network (which has 92% accuracy without any attack), under the strong C&W attack within a certain distortion tolerance, the accuracy of unprotected model drops to less than 10%, the best previous defense technique has 48 % accuracy, while our method still has 86 % prediction accuracy under the same level of attack. Finally, our method is simple and easy to integrate into any neural network.

Original languageEnglish (US)
Title of host publicationComputer Vision – ECCV 2018 - 15th European Conference, 2018, Proceedings
EditorsVittorio Ferrari, Cristian Sminchisescu, Martial Hebert, Yair Weiss
PublisherSpringer Verlag
Pages381-397
Number of pages17
ISBN (Print)9783030012335
DOIs
StatePublished - 2018
Event15th European Conference on Computer Vision, ECCV 2018 - Munich, Germany
Duration: Sep 8 2018Sep 14 2018

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume11211 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference15th European Conference on Computer Vision, ECCV 2018
Country/TerritoryGermany
CityMunich
Period9/8/189/14/18

All Science Journal Classification (ASJC) codes

  • Theoretical Computer Science
  • General Computer Science

Cite this