Bit flipping-based error correcting output code construction for adversarial robustness of neural networks

Abstract

In this paper, we propose a method for constructing error-correcting output codes (ECOCs) based on a codeword bit flipping algorithm to enhance adversarial robustness of neural networks. In the previous work in Verma and Swami (2019), ECOCs are applied to deep neural networks (DNNs) based on the analogy between channel noise and adversarial examples to achieve state-of-the-art adversarial robustness. To improve adversarial robustness, it was proposed in Wan et al. (2022) to optimize the Hamming distance between codewords and employ codeword assignment algorithms. Our study achieves approximately a 8% accuracy improvement on MNIST and CIFAR-10 under adversarial attacks compared to the method proposed in Wan et al. (2022).