Functional Adversarial Attacks

Cassidy Laidlaw, Soheil Feizi

Abstract

We propose functional adversarial attacks, a novel class of threat models for crafting adversarial examples to fool machine learning models. Unlike a standard $\ell_p$-ball threat model, a functional adversarial threat model allows only a single function to be used to perturb input features to produce an adversarial example. For example, a functional adversarial attack applied on colors of an image can change all red pixels simultaneously to light red. Such global uniform changes in images can be less perceptible than perturbing pixels of the image individually. For simplicity, we refer to functional adversarial attacks on image colors as ReColorAdv, which is the main focus of our experiments. We show that functional threat models can be combined with existing additive ($\ell_p$) threat models to generate stronger threat models that allow both small, individual perturbations and large, uniform changes to an input. Moreover, we prove that such combinations encompass perturbations that would not be allowed in either constituent threat model. In practice, ReColorAdv can significantly reduce the accuracy of a ResNet-32 trained on CIFAR-10. Furthermore, to the best of our knowledge, combining ReColorAdv with other attacks leads to the strongest existing attack even after adversarial training. An implementation of ReColorAdv is available at https://github.com/cassidylaidlaw/ReColorAdv .

Introduction

Traditional research into adversarial examples primarily focuses on the $\ell_p$-ball threat model, which assumes an attacker can change any pixel by a small, independent amount. While mathematically convenient, this model fails to capture the nuances of human perception. Humans are highly sensitive to "salt-and-pepper" noise but often overlook global changes, such as shifts in color temperature or lighting. A global change might result in a large $\ell_p$ distance while remaining imperceptible to a person. This paper introduces functional adversarial attacks to exploit this gap, specifically focusing on ReColorAdv, which manipulates image colors globally.

Methodology

The core concept of a functional attack is to apply a single function $f$ to every pixel $x_i$ of an input, such that $\tilde{x}_i = f(x_i)$. This contrasts with traditional additive attacks where a unique noise value $\delta_i$ is added to each pixel.

Conceptual overview of additive, functional, and combined adversarial attacks.

ReColorAdv implements this by mapping input colors to new colors using a discrete 3D grid and trilinear interpolation. To ensure stealthiness, the authors employ perceptual regularization: $F_{diff}$ limits individual color changes, and $F_{smooth}$ ensures similar colors shift consistently. By operating in the perceptually uniform CIELUV color space, the attack can make larger changes in areas less noticeable to humans. Furthermore, functional attacks can be "stacked" with additive and spatial attacks, creating a combined threat model that reaches adversarial regions inaccessible to any single method.

Experimental Results

The authors evaluated ReColorAdv on CIFAR-10 (ResNet-32) and ImageNet (Inception-v4). The results demonstrate devastating effectiveness: ReColorAdv alone reduced CIFAR-10 accuracy from over 92% to just 3.0%.

Examples of ReColorAdv attacks on ImageNet images. Left: original; Middle: adversarial; Right: magnified perturbation.

Even robust models trained with adversarial training are vulnerable. A model's accuracy dropped to 3.6% when faced with a combination of ReColorAdv, spatial, and additive attacks—the lowest reported accuracy for a defended model at the time. Notably, models robust against additive attacks remained highly vulnerable to functional attacks, suggesting that robustness to different perturbation types is "orthogonal."

Conclusion

Functional adversarial attacks represent a powerful new threat model that exploits global input transformations. By combining these with traditional attacks, the authors expose significant vulnerabilities in current state-of-the-art defenses. Future work includes extending this framework to other domains like audio or text and developing unified defenses capable of handling multiple simultaneous threat models.

Summary updated

Abstract

We propose functional adversarial attacks, a novel class of threat models for crafting adversarial examples to fool machine learning models. Unlike a standard $\ell_p$-ball threat model, a functional adversarial threat model allows only a single function to be used to perturb input features to produce an adversarial example. For example, a functional adversarial attack applied on colors of an image can change all red pixels simultaneously to light red. Such global uniform changes in images can be less perceptible than perturbing pixels of the image individually. For simplicity, we refer to functional adversarial attacks on image colors as ReColorAdv, which is the main focus of our experiments. We show that functional threat models can be combined with existing additive ($\ell_p$) threat models to generate stronger threat models that allow both small, individual perturbations and large, uniform changes to an input. Moreover, we prove that such combinations encompass perturbations that would not be allowed in either constituent threat model. In practice, ReColorAdv can significantly reduce the accuracy of a ResNet-32 trained on CIFAR-10. Furthermore, to the best of our knowledge, combining ReColorAdv with other attacks leads to the strongest existing attack even after adversarial training. An implementation of ReColorAdv is available at https://github.com/cassidylaidlaw/ReColorAdv .
BibTeX
@inproceedings{laidlaw2019functional,
  author = {Cassidy Laidlaw and Soheil Feizi},
  title = {{Functional Adversarial Attacks}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2019},
  url = {https://openreview.net/forum?id=T5pjlkbQ6fs},
}
Figure