Multiple Perturbation Attack: Attack Pixelwise Under Different $\ell_p$-norms For Better Adversarial Performance

Ngoc N. Tran, Anh Tuan Bui, Dinh Q. Phung, Trung LearXiv 2022 · OpenReview

Abstract

Adversarial machine learning has been both a major concern and a hot topic recently, especially with the ubiquitous use of deep neural networks in the current landscape. Adversarial attacks and defenses are usually likened to a cat-and-mouse game in which defenders and attackers evolve over the time. On one hand, the goal is to develop strong and robust deep networks that are resistant to malicious actors. On the other hand, in order to achieve that, we need to devise even stronger adversarial attacks to challenge these defense models. Most of existing attacks employs a single $\ell_p$ distance (commonly, $p\in\{1,2,\infty\}$) to define the concept of closeness and performs steepest gradient ascent w.r.t. this $p$-norm to update all pixels in an adversarial example in the same way. These $\ell_p$ attacks each has its own pros and cons; and there is no single attack that can successfully break through defense models that are robust against multiple $\ell_p$ norms simultaneously. Motivated by these observations, we come up with a natural approach: combining various $\ell_p$ gradient projections on a pixel level to achieve a joint adversarial perturbation. Specifically, we learn how to perturb each pixel to maximize the attack performance, while maintaining the overall visual imperceptibility of adversarial examples. Finally, through various experiments with standardized benchmarks, we show that our method outperforms most current strong attacks across state-of-the-art defense mechanisms, while retaining its ability to remain clean visually.

Introduction

Adversarial attacks typically rely on a single $\ell_p$ norm to define "closeness" to the original image. For example, $\ell_\infty$ attacks limit the maximum change to any single pixel, while $\ell_1$ attacks allow large changes to a few pixels. This creates a "cat-and-mouse" game with distinct gaps. Most top-tier defenses are optimized for $\ell_\infty$ threats but remain vulnerable to $\ell_1$ attacks. Furthermore, while $\ell_1$ attacks are mathematically "close," they often produce "salt-and-pepper" noise that is glaringly obvious to the human eye. New "multi-norm" defenses are designed to resist various single-norm attacks simultaneously, and existing attacks struggle to break these models because they only use one type of perturbation logic at a time.

The Multiple Perturbation Attack (MPA) moves beyond the "one-norm-fits-all" approach by optimizing perturbations at the pixel level. By dynamically choosing the most effective $\ell_p$ norm ($p \in \{1, 2, \infty\}$) for every single pixel, MPA achieves significantly higher attack success rates against state-of-the-art defenses while remaining visually stealthier than traditional $\ell_1$ attacks.

Methodology

The core innovation of MPA is pixel-wise norm selection. Instead of applying one norm to the entire image, the algorithm decides which norm is most effective for each individual pixel.

  1. Steepest Ascent Directions: For every iteration, the algorithm calculates the gradient of the loss function and determines the "steepest ascent" direction for $\ell_1, \ell_2,$ and $\ell_\infty$ norms.
  2. Temperature Softmax Mixing: To make the selection process differentiable, a "soft" coefficient tensor $c$ is used. A temperature softmax is applied to these coefficients, allowing the model to "learn" which norm maximizes the loss for each pixel during an inner optimization loop.
  3. Hard Decision & Custom Projection: After the inner loop, the "soft" weights are converted into a "hard" choice (picking the single best norm per pixel). A custom projection operator ensures that the sub-regions of the image perturbed by different norms stay within their respective "imperceptibility" budgets ($\epsilon$-balls).
Visual comparison of MPA against traditional single-norm attacks across different datasets.

Experiments and Results

The researchers evaluated MPA on standard computer vision benchmarks including CIFAR-10, CIFAR-100, and ImageNet. MPA demonstrates that different pixels in an image are vulnerable to different types of noise. Forcing a uniform perturbation strategy across the whole image is suboptimal for the attacker.

Performance Against State-of-the-Art Defenses

On ImageNet, MPA outperformed all single-norm attacks (including the gold-standard AutoAttack) across multiple defense models. As shown in the table below, it reduced the robust accuracy of the Debenedetti et al. model to 53.46%, whereas the best previous attack only reached 55.40%.

Robust accuracy comparison on ImageNet, showing MPA's superior performance against state-of-the-art defenses.

MPA was particularly devastating against multi-norm models. On CIFAR-10, it was highly effective against the MSD (multi-norm) defense, achieving a robust accuracy of 26.0%—a massive 23% drop compared to the next best attack (AutoAttack-$\ell_1$).

Visual Quality

Quantitative metrics and qualitative visualizations show that MPA results are nearly indistinguishable from the original images. It acts as a bridge, capturing the high potency of $\ell_1$ attacks but distributing that energy in a way that mimics the visual "cleanliness" of $\ell_2$ and $\ell_\infty$ attacks, avoiding the "weird pixels" common in pure $\ell_1$ attacks.

Conclusion

While MPA shows superior potency, it does come with computational overhead, requiring roughly three times the memory of a standard single-norm attack. Future work may explore better coefficient initialization schemes and investigate whether training models against MPA could lead to even more robust "universal" defenses. Overall, MPA demonstrates that pixel-level flexibility is a powerful tool for creating stronger, more natural adversarial examples.

Summary updated

Abstract

Adversarial machine learning has been both a major concern and a hot topic recently, especially with the ubiquitous use of deep neural networks in the current landscape. Adversarial attacks and defenses are usually likened to a cat-and-mouse game in which defenders and attackers evolve over the time. On one hand, the goal is to develop strong and robust deep networks that are resistant to malicious actors. On the other hand, in order to achieve that, we need to devise even stronger adversarial attacks to challenge these defense models. Most of existing attacks employs a single $\ell_p$ distance (commonly, $p\in\{1,2,\infty\}$) to define the concept of closeness and performs steepest gradient ascent w.r.t. this $p$-norm to update all pixels in an adversarial example in the same way. These $\ell_p$ attacks each has its own pros and cons; and there is no single attack that can successfully break through defense models that are robust against multiple $\ell_p$ norms simultaneously. Motivated by these observations, we come up with a natural approach: combining various $\ell_p$ gradient projections on a pixel level to achieve a joint adversarial perturbation. Specifically, we learn how to perturb each pixel to maximize the attack performance, while maintaining the overall visual imperceptibility of adversarial examples. Finally, through various experiments with standardized benchmarks, we show that our method outperforms most current strong attacks across state-of-the-art defense mechanisms, while retaining its ability to remain clean visually.
BibTeX
@misc{tran2022multiple,
  author = {Ngoc N. Tran and Anh Tuan Bui and Dinh Q. Phung and Trung Le},
  title = {{Multiple Perturbation Attack: Attack Pixelwise Under Different $\textbackslash\{\}ell\_p$-norms For Better Adversarial Performance}},
  year = {2022},
  eprint = {2212.03069},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2212.03069},
}
Figure