One pixel attack for fooling deep neural networks

TL;DR

This paper introduces a novel one-pixel attack method using differential evolution to fool deep neural networks, revealing their vulnerability even with minimal input modifications.

Contribution

The paper presents a black-box one-pixel attack leveraging differential evolution, demonstrating its effectiveness across multiple datasets and highlighting a new low-dimensional adversarial attack approach.

Findings

67.97% of CIFAR-10 images can be fooled with one pixel change

16.04% of ImageNet images can be fooled with one pixel change

The attack requires minimal adversarial information and is highly effective

Abstract

Recent research has revealed that the output of Deep Neural Networks (DNN) can be easily altered by adding relatively small perturbations to the input vector. In this paper, we analyze an attack in an extremely limited scenario where only one pixel can be modified. For that we propose a novel method for generating one-pixel adversarial perturbations based on differential evolution (DE). It requires less adversarial information (a black-box attack) and can fool more types of networks due to the inherent features of DE. The results show that 67.97% of the natural images in Kaggle CIFAR-10 test dataset and 16.04% of the ImageNet (ILSVRC 2012) test images can be perturbed to at least one target class by modifying just one pixel with 74.03% and 22.91% confidence on average. We also show the same vulnerability on the original CIFAR-10 dataset. Thus, the proposed attack explores a different take on adversarial machine learning in an extreme limited scenario, showing that current DNNs are also vulnerable to such low dimension attacks. Besides, we also illustrate an important application of DE (or broadly speaking, evolutionary computation) in the domain of adversarial machine learning: creating tools that can effectively generate low-cost adversarial attacks against neural networks for evaluating robustness.