基于Fire Module卷积神经网络的手写变造数字检测-豆柴文库

您所在位置：网站首页 / 基于Fire Module卷积神经网络的手写变造数字检测.docx / 文档详情

在线预览结束，喜欢就下载吧，查找使用更方便

5 金币

下载文档

/ 2

下载提示

如果您无法下载资料，请参考说明：

1、部分资料下载需要金币，请确保您的账户上有足够的金币

2、已购买过的文档，再次下载不重复扣费

3、资料包下载后请先用软件解压，在使用对应软件打开

文本预览

基于FireModule卷积神经网络的手写变造数字检测
Title:HandwrittenAdversarialDigitDetectionusingConvolutionalNeuralNetworkswithFireModule
Abstract:
Handwrittendigitrecognitionhaswitnessedsignificantadvancementsinrecentyears,withconvolutionalneuralnetworks(CNNs)emergingasthestate-of-the-artsolution.However,thesemodelsarevulnerabletoadversarialattacks,whereimperceptibleperturbationsareaddedtotheinputimage,causingmisclassification.Inthispaper,weproposeanovelapproachfordetectingadversarialattacksonhandwrittendigitsusingCNNswithFireModules.TheFireModuleleveragestheadvantagesofbothdepthwiseseparableconvolutionsandsqueeze-and-excitationoperationstoenhancethenetwork'srobustnessanddiscriminativeability.Experimentalresultsdemonstratetheeffectivenessofourproposedmethodindetectingadversarialattackswhilemaintaininghighaccuracyonlegitimatehandwrittendigitrecognitiontasks.
Keywords:Handwrittendigitrecognition,Adversarialattacks,ConvolutionalNeuralNetworks,FireModule,Robustness
1.Introduction
Handwrittendigitrecognitionplaysacrucialroleinvariousapplications,suchaspostalserviceautomation,bankcheckprocessing,anddigitclassificationforremotesensing.CNNshaveachievedremarkablesuccessinthisfieldbyeffectivelyextractingdiscriminativefeaturesfromdigitimages.However,recentresearchhasshownthatCNNsaresusceptibletoadversarialattacks,whereanattackerintentionallymanipulatestheinputimagetocausemisclassificationatinferencetime.
2.AdversarialAttacksonHandwrittenDigits
Adversarialattacksonhandwrittendigitsinvolvethecreationofimperceptibleperturbationstotheinputimage,leadingtothemisclassificationofthetargetdigit.Theseperturbationsarecarefullydesignedtodeceivethenetworkwhileremainingvisuallysimilartotheoriginalimage.CommonattackmethodsincludetheFastGradientSignMethod(FGSM)andtheProjectedGradientDescent(PGD)algorithm.TheseattacksposeasignificantchallengetothedeploymentofCNNmodelsinreal-worldapplications.
3.FireModule:EnhancingRobustnessinCNNs
TheFireModule,originallyproposedintheSqueezeNetarchitecture,isacompactandefficientm