1 Universal Adversarial Perturbations Efﬁciency on a small image dataset

2025-04-28 0 0 3.42MB 5 页 10玖币

侵权投诉

Universal Adversarial Perturbations: Efﬁciency

on a small image dataset

Waris Radji*

Index Terms—Computer vision, Image classiﬁcation, Deep neural networks, Adversarial attack.

ABSTRACT

Although neural networks perform very well on the image classi-

ﬁcation task, they are still vulnerable to adversarial perturbations

that can fool a neural network without visibly changing an input

image. A paper has shown the existence of Universal Adversarial

Perturbations which when added to any image will fool the

neural network with a very high probability. In this paper we

will try to reproduce the experience of the Universal Adversarial

Perturbations paper, but on a smaller neural network architecture

and training set, in order to be able to study the efﬁciency of the

computed perturbation.

1 INTRODUCTION

As part of my engineering school’s introductory research

program, I aim to reproduce a scientiﬁc experiment. I

chose to reproduce the universal adversarial perturbations

paper experiments introduced by Seyed-Mohsen Moosavi-

Dezfooli et al accepted at IEEE Conference on Computer

Vision and Pattern Recognition (CVPR) in 2017[1].

In image classiﬁcation, an adversarial perturbation is a

vector that when added to an image cause the network out-

put to change drastically with making quasi-imperceptible

changes to the input image. Universal adversarial pertur-

bations (UAPs) correspond to a single vector that will fool

the neural network on any image with very high probability

(see Fig.1). These perturbations have many applications in

the real world, for example in CAPTCHA tests when a user

has to associate an image with a label or in steganography

to hide information in images[2]. They can also be used

for hostile purposes, knowing that they are very difﬁcult

to detect with the naked eye.

The reference paper [1] has already demonstrated the

existence and efﬁciency of UAPs, on large neural net-

work architecture (CaffeNet, VGG-F, VGG-16, VGG-19,

GoogLeNet, and ResNet-152) and with the very large im-

age dataset of ImageNet Large Scale Visual Recognition

Challenge 2012 (ILSVRC2012). In this paper we will try to

reproduce some of their results, with VGG-11 [5] neural

network architecture which is smaller than those studied

in the paper and on the Visual Object Classes Challenge

2012 (VOC2012) [4] dataset which is 1,000 times smaller than

ILSVRC2012:

*Student in the introductory research program of Bordeaux Institute of

Technology - ENSEIRB-MATMECA School of Engineering.

ick

ail

iza

Balloon

Lycaenid

Tibetan masti

ril

lag

Fac

e p

Labrador

hih

Labrador

Tibetan masti

Brabancon grion

Border terrier

Fig. 1: This ﬁgure is taken from the reference article. When

added to a natural image, a universal perturbation image

causes the image to be misclassiﬁed by the deep neural

network with high probability. Left images: Original natural

images. The labels are shown on top of each arrow. Cen-

tral image: Universal perturbation. Right images: Perturbed

images.

•We will train a neural network;

•We will compute the UAP with parameters quasi-

similar to the paper one;

•We will observe dominant labels founded by the

algorithm;

•We will compare the computed UAP to other pertur-

bations.

All the code that allowed me to achieve my experiments

can be found on Github.

arXiv:2210.04591v1 [cs.CV] 10 Oct 2022

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

1UniversalAdversarialPerturbations:EfciencyonasmallimagedatasetWarisRadji*IndexTermsComputervision,Imageclassication,Deepneuralnetworks,Adversarialattack.FABSTRACTAlthoughneuralnetworksperformverywellontheimageclassi-cationtask,theyarestillvulnerabletoadversarialperturbationsthatcanfoolaneuralne...

展开>> 收起<<

1 Universal Adversarial Perturbations Efﬁciency on a small image dataset.pdf

共5页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

1 Universal Adversarial Perturbations Efﬁciency on a small image dataset

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: