ADef: an Iterative Algorithm to Construct Adversarial Deformations

TL;DR

This paper introduces ADef, an iterative algorithm that creates adversarial attacks by applying small deformations to images, enhancing understanding of neural network vulnerabilities.

Contribution

The paper presents a novel deformation-based adversarial attack method and demonstrates its effectiveness on MNIST and ImageNet datasets.

Findings

ADef successfully generates adversarial deformations.

Deformations can fool neural networks like Inception-v3 and ResNet-101.

The method reveals new insights into neural network stability.

Abstract

While deep neural networks have proven to be a powerful tool for many recognition and classification tasks, their stability properties are still not well understood. In the past, image classifiers have been shown to be vulnerable to so-called adversarial attacks, which are created by additively perturbing the correctly classified image. In this paper, we propose the ADef algorithm to construct a different kind of adversarial attack created by iteratively applying small deformations to the image, found through a gradient descent step. We demonstrate our results on MNIST with convolutional neural networks and on ImageNet with Inception-v3 and ResNet-101.