[Paper Review] Deep Neural Network Ensembles against Deception: Ensemble Diversity, Accuracy and Robustness

Ling Liu, Wenqi Wei+5 more|arXiv (Cornell University)|Aug 29, 2019

Adversarial Robustness in Machine Learning23 references7 citations

TL;DR

This paper proposes a deep neural network (DNN) ensemble framework that enhances accuracy and robustness against adversarial attacks by optimizing ensemble diversity through multiple diversity measures and consensus strategies. By strategically combining diverse DNN models via voting or learned fusion, the ensemble achieves higher accuracy and resists deception from adversarial perturbations.

ABSTRACT

Ensemble learning is a methodology that integrates multiple DNN learners for improving prediction performance of individual learners. Diversity is greater when the errors of the ensemble prediction is more uniformly distributed. Greater diversity is highly correlated with the increase in ensemble accuracy. Another attractive property of diversity optimized ensemble learning is its robustness against deception: an adversarial perturbation attack can mislead one DNN model to misclassify but may not fool other ensemble DNN members consistently. In this paper we first give an overview of the concept of ensemble diversity and examine the three types of ensemble diversity in the context of DNN classifiers. We then describe a set of ensemble diversity measures, a suite of algorithms for creating diversity ensembles and for performing ensemble consensus (voted or learned) for generating high accuracy ensemble output by strategically combining outputs of individual members. This paper concludes with a discussion on a set of open issues in quantifying ensemble diversity for robust deep learning.

Motivation & Objective

• To investigate how ensemble diversity improves the accuracy and robustness of deep neural network classifiers.
• To identify and quantify three types of diversity in DNN ensembles for improved generalization and resistance to adversarial attacks.
• To develop a suite of algorithms that generate diverse DNN ensembles and effective consensus mechanisms for high-accuracy predictions.
• To address open challenges in measuring and leveraging ensemble diversity for robust deep learning.

Proposed method

• Categorizes and analyzes three types of ensemble diversity in DNN classifiers: error diversity, output diversity, and parameter diversity.
• Introduces a set of quantitative diversity measures to evaluate and guide the creation of diverse DNN ensembles.
• Employs training strategies such as data shuffling, model weight initialization, and data augmentation to promote diversity among ensemble members.
• Applies both voted and learned consensus methods to combine predictions from individual DNN models for final output.
• Uses adversarial training and perturbation testing to evaluate robustness against deception.
• Proposes a framework for systematic evaluation of diversity metrics in relation to ensemble performance and adversarial robustness.

Experimental results

Research questions

• RQ1 How does ensemble diversity correlate with improved accuracy and robustness in DNN classifiers?
• RQ2 What are the three primary types of diversity in DNN ensembles, and how can they be measured and optimized?
• RQ3 To what extent can diversity-based ensemble learning resist adversarial perturbation attacks compared to single models?
• RQ4 How do different consensus strategies (voting vs. learned fusion) affect ensemble performance and robustness?
• RQ5 What are the key open challenges in quantifying and leveraging ensemble diversity for robust deep learning?

Key findings

• Greater ensemble diversity leads to more uniformly distributed errors, which significantly increases overall ensemble accuracy.
• Diversity-optimized ensembles demonstrate superior robustness against adversarial attacks, as perturbations that fool one model often fail to mislead the entire ensemble.
• The use of multiple diversity measures enables systematic identification and enhancement of complementary DNN models within the ensemble.
• Learned consensus methods outperform simple voting in some settings by better capturing inter-model relationships and improving final prediction accuracy.
• The paper identifies critical open issues in defining and measuring diversity in DNN ensembles, particularly in relation to adversarial robustness.
• Empirical results show that diverse ensembles maintain high accuracy even under targeted adversarial attacks, highlighting their resilience to deception.