Construction of Error Correcting Output Codes for Robust Deep Neural Networks Based on Label Grouping Scheme

Abstract

Error-Correcting Output Codes (ECOCs) have been proposed to construct multi-class classifiers using simple binary classifiers. Recently, the principle of ECOCs has been employed for improving the robustness of deep classifiers. In this paper, a novel ECOC framework is developed by presenting a novel label grouping and code-construction method. The proposed label grouping is based on linear discriminant analysis (LDA) similarity. Via simulations, it is demonstrated that deep classifiers trained with the proposed ECOC yield better classification performance on pure data and better adversarial robustness than the state-of-the-art deep neural classifiers using ECOCs.