AAT: An Efficient Adaptive Adversarial Training Algorithm

Abstract

Adversarial training is one of the most promising methods to improve the model's robustness, while the expensive training cost keeps a huge problem for this method. Recent researchers have made great effort to improve its performance by reducing the inner adversarial sample construction cost. Their works have alleviated this problem to some extent while the overall performance is still expensive and not interpretable. In this work, we propose AAT (Adaptive Adversarial Training) algorithm utilizing the inherent relationship between the model's robustness and the effects of the adversarial samples to accelerate the overall performance. Our method offers more interpretable robustness improvement while achieving higher efficiency than the state-of-the-art works on standard datasets. We have reduced more than 56% training time than traditional adversarial training on CIFAR10.