Should Adversarial Attacks Use Pixel p-Norm?

Ayon Sen, Xiaojin Zhu, Liam Marshall, Robert D. NowakarXiv 2019 · arXiv

Abstract

Adversarial machine learning research typically uses pixel $p$-norms to quantify the "imperceptibility" of perturbations. This paper presents a behavioral study involving human subjects to test whether these mathematical metrics align with visual perception. By measuring the "just noticeable difference" (JND) across various perturbation types, the authors demonstrate that standard norms ($L_0, L_1, L_2, L_\infty$) and perceptual metrics like SSIM are poor proxies for human vision. Notably, the study identifies the pixel 3-norm ($L_3$) as a more consistent, though still imperfect, approximation of human perceptual thresholds.

The Proxy Problem in Adversarial ML

In adversarial machine learning, an attack is considered successful if it alters a model's output while remaining invisible to a human observer. To formalize this, researchers rely on the "Central Hypothesis": that there exists a specific $p$-norm and a radius $r$ such that any perturbation within that radius is imperceptible. Consequently, $L_p$ distances (most commonly $L_2$ and $L_\infty$) have become the standard for defining the threat models of adversarial attacks. However, there is little empirical evidence that these mathematical distances actually correspond to the complexities of human visual processing.

Experimental Design: Measuring the JND

To test the validity of $p$-norms, the researchers conducted a behavioral experiment with 42 reliable participants from Amazon Mechanical Turk. The core task was a "Just Noticeable Difference" (JND) test: subjects were shown an original image ($x_0$) and asked to manually adjust the intensity of a perturbation ($v$) until they could just barely detect a change.

The study utilized three iconic images—a panda, a macaw, and a cat—and tested ten distinct perturbation directions. These included handcrafted geometric changes (e.g., red dots, centered boxes, changes to animal eyes) and adversarial directions generated by FGSM and PGD attacks.

Examples of the ten perturbation directions (a-j) applied to the panda, macaw, and cat datasets, ranging from sparse dots to adversarial noise.

Evaluating Perceptual Metrics

The study used the Kolmogorov-Smirnov (KS) test to evaluate if JND values across different directions followed the same distribution for a given norm. If a norm were a perfect proxy for perception, the JND should occur at a consistent value regardless of the perturbation's shape.

The results decisively rejected all standard $p$-norms, as well as Structural Similarity Index (SSIM), Earth Mover’s Distance (EMD), and Deep Neural Network (DNN) embedding distances. A key finding was that context matters: humans detect changes to semantically important features (like a macaw's eye) much faster than random noise, even when the $p$-norm is identical.

The 3-Norm Surprise

By analyzing the "variability of fit" (the standard deviation of normalized JND values), the authors discovered that the pixel 3-norm ($L_3$) consistently outperformed more common metrics. As shown in the figure below, the 3-norm achieves the lowest variability across all three image sets, suggesting it is the "least bad" mathematical approximation of human visual data among the tested norms.

Variability of fit for various p-norms and perceptual metrics. The pixel 3-norm shows the lowest standard deviation, indicating it is most consistent with human JND data.

Conclusion

The findings suggest that the adversarial machine learning community's reliance on $L_2$ and $L_\infty$ norms may be misguided. These metrics fail to capture how humans perceive perturbations, potentially leading to the study of "adversarial" examples that are easily spotted or the neglect of stealthy attacks with large mathematical norms. While the 3-norm offers a better fit, the failure of even sophisticated metrics like SSIM and DNN embeddings underscores the ongoing challenge of modeling human vision mathematically. Future work must look beyond simple pixel-wise distances toward metrics that account for semantic importance and spatial context.

Summary updated

BibTeX
@misc{sen2019should,
  author = {Ayon Sen and Xiaojin Zhu and Liam Marshall and Robert D. Nowak},
  title = {{Should Adversarial Attacks Use Pixel p-Norm?}},
  year = {2019},
  eprint = {1906.02439},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1906.02439},
}
Figure