Towards fast and effective single-step adversarial training
Recently, Wong et al. (2020) showed adversarial training with single-step FGSM leads to a characteristic failure mode named catastrophic overfitting (CO), in which a model becomes suddenly vulnerable to multi-step attacks. Moreover, they showed adding a random perturbation prior to FGSM (RS-FGSM) seemed to be sufficient to prevent CO. However, Andriushchenko & Flammarion (2020) observed that RS-FGSM still leads to CO for larger perturbations and argue that the only contribution of the random step is to reduce the magnitude of the attacks. They suggest a regularizer (GradAlign) that avoids CO but is significantly more expensive than RS-FGSM. In this work, we methodically revisit the role of noise and clipping in single-step adversarial training. Contrary to previous intuitions, we find that not clipping the perturbation around the clean sample and using a stronger noise is highly effective in avoiding CO for large perturbation radii, despite leading to an increase in the magnitude of the attacks. Based on these observations, we propose a method called Noise-FGSM (N-FGSM), which attacks noise-augmented samples directly using a single-step. Empirical analyses on a large suite of experiments show that N-FGSM is able to match or surpass the performance of GradAlign while achieving a 3x speed-up.
PDF Abstract
CIFAR-10
CIFAR-100
SVHN
