Adversarial Examples for Semantic Image Segmentation

TL;DR

This paper investigates how adversarial perturbations can be applied to semantic image segmentation, demonstrating that existing attacks can cause widespread misclassification of specific classes with minimal perceptible changes.

Contribution

It extends adversarial attack analysis from image classification to semantic segmentation, showing transferability and effectiveness of imperceptible perturbations in this context.

Findings

Adversarial attacks can transfer from classification to segmentation tasks.

Imperceptible perturbations can cause misclassification of almost all pixels of a target class.

The network's predictions outside the targeted class remain nearly unchanged.

Abstract

Machine learning methods in general and Deep Neural Networks in particular have shown to be vulnerable to adversarial perturbations. So far this phenomenon has mainly been studied in the context of whole-image classification. In this contribution, we analyse how adversarial perturbations can affect the task of semantic segmentation. We show how existing adversarial attackers can be transferred to this task and that it is possible to create imperceptible adversarial perturbations that lead a deep network to misclassify almost all pixels of a chosen class while leaving network prediction nearly unchanged outside this class.