Fast Adversarial Training with Noise Augmentation: A Unified Perspective on RandStart and GradAlign

Chaoning Zhang, Kang Zhang, Axi Niu, Chenshuang Zhang, Jiu Feng, Chang D. Yoo, In So Kweon

Abstract

PGD-based and FGSM-based are two popular adversarial training (AT) approaches for obtaining adversarially robust models. Compared with PGD-based AT, FGSM-based one is significantly faster but fails with catastrophic overfitting (CO). For mitigating CO in such Fast AT, there are two popular existing strategies: random start (RandStart) and Gradient Alignment (GradAlign). The former works only for a relatively small perturbation 8/255 with the $l_\infty$ constraint, and GradAlign improves it by extending the perturbation size to 16/255 but at the cost of being 3 to 4 times slower. How to avoid CO in Fast AT for a large perturbation size but without increasing the computation overhead remains as an unsolved issue, for which our work provides a frustratingly simple (yet effective) solution. Specifically, our solution lies in just noise augmentation (NoiseAug) which is a non-trivial byproduct of simplifying GradAlign. By simplifying GradAlign we have two findings: (i) aligning logit instead of gradient in GradAlign requires half the training time but achieves higher performance than GradAlign; (ii) the alignment operation can also be removed by only keeping noise augmentation (NoiseAug). Simplified from GradAlign, our NoiseAug has a surprising resemblance with RandStart except that we inject noise on the image instead of perturbation. To understand why injecting noise to input prevents CO, we verify that this is caused not by data augmentation effect but by improved local linearity. We provide an intuitive explanation for why NoiseAug improves local linearity without explicit regularization. Extensive results demonstrate that our NoiseAug achieves SOTA results in FGSM AT.

Introduction

Deep neural networks are notoriously vulnerable to adversarial examples—inputs with tiny, invisible perturbations designed to cause misclassification. To defend against them, researchers use Adversarial Training (AT), where the model is trained on these malicious examples. The gold standard is PGD-AT (Projected Gradient Descent), but it is computationally expensive because it requires multiple iterations to generate each adversarial example.

A faster alternative is FGSM-AT (Fast Gradient Sign Method), which uses only a single step. However, FGSM-AT often suffers from Catastrophic Overfitting (CO): a failure mode where the model suddenly loses all robustness against strong attacks during training, even if it looks "robust" against the weak FGSM attack it was trained on. Previous fixes like RandStart only work for small perturbations ($\epsilon=8/255$), while GradAlign works for larger ones ($\epsilon=16/255$) but makes training 3 to 4 times slower. This paper introduces NoiseAug, a method as fast as RandStart but as effective as GradAlign.

Methodology: From GradAlign to NoiseAug

The authors arrived at their solution by systematically simplifying GradAlign, a method that regularizes the model to keep gradients of clean and noisy images similar.

  1. LogitAlign: They first found that aligning the outputs (logits) of clean and noisy images instead of their gradients was twice as fast and slightly more effective.
  2. NoiseAug: They then discovered that the alignment operation itself wasn't strictly necessary. Simply augmenting the training images with noise before the FGSM step was enough to prevent CO.
The simplification path from GradAlign to LogitAlign and finally NoiseAug, showing the structural resemblance between NoiseAug and RandStart.

In the NoiseAug algorithm, for every training image $x$, the method adds random noise ($x' = x + \eta$), generates an FGSM perturbation $\delta$ based on $x'$, and updates the model using the adversarial example $x' + \delta$. This disentangles the "noise" from the "perturbation," whereas previous methods like RandStart treated the noise as the starting point of the perturbation itself, limiting its magnitude to the perturbation budget $\epsilon$.

Why NoiseAug Works: Local Linearity

The authors prove that NoiseAug prevents CO not because it acts as a standard data augmenter (like Mixup or Cutout, which fail to prevent CO), but because it improves the local linearity of the model's loss landscape. By looking at feature maps, they found that standard FGSM-AT develops "Laplace-like" filters that are hyper-sensitive to noise. NoiseAug prevents these filters from forming, leading to a smoother, more robust model.

Cosine similarity of gradients across different layers, demonstrating that NoiseAug maintains high alignment comparable to explicit regularization methods like GradAlign.

As shown in the gradient analysis, NoiseAug maintains high cosine similarity between gradients of clean and noisy inputs throughout the network layers. This suggests that NoiseAug implicitly achieves the goal of GradAlign without the explicit, computationally expensive regularization term.

Experimental Results

The researchers evaluated NoiseAug on CIFAR-10 and SVHN using PreAct ResNet18 and WideResNet28-10 models. Key findings include:

  • Efficiency: NoiseAug adds zero computational overhead compared to standard RandStart and is roughly 4x faster than GradAlign.
  • Robustness at Scale: At a large perturbation budget of $\epsilon=16/255$ on CIFAR-10, where RandStart fails completely (0% robustness), NoiseAug maintains 29.4% robustness against a strong PGD-50 attack.
  • SOTA Performance: On WideResNet28-10, NoiseAug achieved 47.24% AutoAttack robustness, outperforming other "Fast AT" variants like YOPO, Free-AT, and GradAlign.

Conclusion

NoiseAug provides a simple yet powerful solution to the catastrophic overfitting problem in fast adversarial training. By disentangling noise augmentation from the perturbation budget, it achieves the stability of complex regularization methods at the speed of the simplest single-step attacks. While it does not yet fully match the absolute robustness of multi-step PGD training, it represents a significant step forward in making robust training practical for large-scale applications.

Summary updated

Abstract

PGD-based and FGSM-based are two popular adversarial training (AT) approaches for obtaining adversarially robust models. Compared with PGD-based AT, FGSM-based one is significantly faster but fails with catastrophic overfitting (CO). For mitigating CO in such Fast AT, there are two popular existing strategies: random start (RandStart) and Gradient Alignment (GradAlign). The former works only for a relatively small perturbation 8/255 with the l_\infty constraint, and GradAlign improves it by extending the perturbation size to 16/255 (with the l_\infty constraint) but at the cost of being 3 to 4 times slower. How to avoid CO in Fast AT for a large perturbation size but without increasing the computation overhead remains as an unsolved issue, for which our work provides a frustratingly simple (yet effective) solution. Specifically, our solution lies in just noise augmentation (NoiseAug) which is a non-trivial byproduct of simplifying GradAlign. By simplifying GradAlign we have two findings: (i) aligning logit instead of gradient in GradAlign requires half the training time but achieves higher performance than GradAlign; (ii) the alignment operation can also be removed by only keeping noise augmentation (NoiseAug). Simplified from GradAlign, our NoiseAug has a surprising resemblance with RandStart except that we inject noise on the image instead of perturbation. To understand why injecting noise to input prevents CO, we verify that this is caused not by data augmentation effect (inject noise on image) but by improved local linearity. We provide an intuitive explanation for why NoiseAug improves local linearity without explicit regularization. Extensive results demonstrate that our NoiseAug achieves SOTA results in FGSM AT. The code will be released after accepted.
BibTeX
@misc{zhang2022fast,
  author = {Chaoning Zhang and Kang Zhang and Axi Niu and Chenshuang Zhang and Jiu Feng and Chang D. Yoo and In So Kweon},
  title = {{Fast Adversarial Training with Noise Augmentation: A Unified Perspective on RandStart and GradAlign}},
  year = {2022},
  eprint = {2202.05488},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2202.05488},
}
Figure