Deep Neural Networks (DNNs) have achieved remarkable performance on a variety of applications but are extremely vulnerable to adversarial perturbation. To address this issue, various defense methods have been proposed to enhance model robustness. Unfortunately, the most representative and promising methods, such as adversarial training and its variants, usually degrade model accuracy on benign samples, limiting practical utility. This indicates that it is difficult to extract both robust and accurate features using a single network under certain conditions, such as limited training data, resulting in a trade-off between accuracy and robustness. To tackle this problem, we propose an Adversarial Feature Stacking (AFS) model that can jointly take advantage of features with varied levels of robustness and accuracy, thus significantly alleviating the aforementioned trade-off. Specifically, we adopt multiple networks adversarially trained with different perturbation budgets to extract either more robust features or more accurate features. These features are then fused by a learnable merger to give final predictions. We evaluate the AFS model on CIFAR-10 and CIFAR-100 datasets with strong adaptive attack methods, which significantly advances the state-of-the-art in terms of the trade-off. Without extra training data, the AFS model achieves a benign accuracy improvement of 6% on CIFAR-10 and 9% on CIFAR-100 with comparable or even stronger robustness than the state-of-the-art adversarial training methods. This work demonstrates the feasibility to obtain both accurate and robust models under the circumstances of limited training data.
Download