Active machine learning approach to adversarial training improves trade-off between natural accuracy and adversarial robustness

Abstract

Abstract Understanding our world which is open and diverse requires foundation models that generalize well while trustworthy. Adversarial training has been considered to be one of the most effective strategies to achieve robust learning systems, yet adversarial training methods have exhibited a trade-off between generalization accuracy and robustness. Motivated by the active machine learning approach to adversarial training, we introduce novel data generation technique acquires adversarial examples based on classification margin criterion, addressing key trade-off between generalization accuracy and adversarial robustness which remains a fundamental challenge in adversarial training. Here, we provide theoretical contribution that sheds light on the properties of the approach which is expected to be beneficial. Likewise, we empirically demonstrate that the proposed method achieves significant improvements in accuracy and robustness with WRN34-10 and ResNet-18 on CIFAR-10, CIFAR-100, SVHN, and TinyImagenet-200. This Article contributes to the broader scientific literature on adversarial training, generalization theory, and robust machine learning.