Multi-attacks: Many images + the same adversarial attack → many target labels

Stanislav Fort

Introduction

In traditional adversarial machine learning, an attacker typically finds a tiny, human-imperceptible change to a single image to trick a model. This research explores a more ambitious question: Can a single "master" perturbation be added to an entire batch of different images to change all of them into different target classes simultaneously? These "multi-attacks" reveal that the decision boundaries of deep learning models are far more complex and redundant than previously realized, posing a significant challenge for exhaustive defense strategies.

Probing the Geometry of Pixel Space

The methodology for generating multi-attacks is strikingly simple. Instead of optimizing a perturbation for one image, the researchers use the Adam optimizer to minimize a joint cross-entropy loss across a whole batch. The goal is to find a single noise pattern $P$ such that for every image $X_i$ in a batch, the model predicts a pre-selected target class $c^*_i$.

A simultaneous multi-attack on 5 images and a 2D section of pixel space spelling 'AGI'.

To further demonstrate the flexibility of these boundaries, the authors optimized two perturbations ($P_x$ and $P_y$) to define a 2D "plane" in pixel space. By ensuring specific coordinates are classified as specific classes, they were able to "paint" shapes and words—such as spelling "AGI" or outlining a tortoise—using the model's own classification logic.

Key Insights: The $10^{100}$ Problem

The study suggests that neural networks are surrounded by a chaotic "foam" of different classification regions. It is estimated that there are roughly $10^{O(100)}$ distinct high-confidence regions around any given image. This staggering redundancy implies that defense strategies relying on mapping all possible adversarial examples are likely impractical.

Resolution as a Vulnerability

A critical finding is that higher-resolution images are easier to attack. As the number of pixels increases, the "room" to find a successful multi-attack grows. The number of successfully attackable images scales linearly with the logarithm of the image resolution.

Impact of image resolution and optimization steps on the success rate of multi-attacks.

Memorization vs. Generalization

Interestingly, models trained on random labels—which are forced to memorize noise rather than learn semantic features—are even more susceptible to multi-attacks. This suggests that "messier" internal representations lead to more fragmented decision boundaries. Furthermore, some adversarial directions are surprisingly stable; the researchers found "lines" in pixel space where an image remains misclassified even as the intensity of the attack increases.

Results and Limitations

Using standard benchmarks like ImageNet (ResNet50) and CIFAR-10, the researchers achieved 100% success in changing the labels of batches of up to 160 images simultaneously with a single perturbation at $224 \times 224$ resolution. While ensembling multiple models provides some resistance, it does not eliminate the vulnerability; larger ensembles merely require more optimization steps or stronger perturbations to crack.

One limitation is the perturbation magnitude; some experiments used $L_\infty$ norms (e.g., 120/255) higher than standard "imperceptible" thresholds. However, the required strength decreases as resolution increases. Future work may involve studying "resistant" images to find clues for building more robust architectures and developing a rigorous mathematical theory to characterize these high-dimensional class boundaries.

Summary updated

Abstract

We show that we can easily design a single adversarial perturbation $P$ that changes the class of $n$ images $X_1,X_2,\dots,X_n$ from their original, unperturbed classes $c_1, c_2,\dots,c_n$ to desired (not necessarily all the same) classes $c^*_1,c^*_2,\dots,c^*_n$ for up to hundreds of images and target classes at once. We call these \textit{multi-attacks}. Characterizing the maximum $n$ we can achieve under different conditions such as image resolution, we estimate the number of regions of high class confidence around a particular image in the space of pixels to be around $10^{\mathcal{O}(100)}$, posing a significant problem for exhaustive defense strategies. We show several immediate consequences of this: adversarial attacks that change the resulting class based on their intensity, and scale-independent adversarial examples. To demonstrate the redundancy and richness of class decision in the pixel space, we look for its two-dimensional sections that trace images and spell words using particular classes. We also show that ensembling reduces susceptibility to multi-attacks, and that classifiers trained on random labels are more susceptible. Our code is available on GitHub.
BibTeX
@misc{fort2023multi,
  author = {Stanislav Fort},
  title = {{Multi-attacks: Many images + the same adversarial attack → many target labels}},
  year = {2023},
  eprint = {2308.03792},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2308.03792},
}
Figure