Adversarial Examples Detection in Deep Networks with Convolutional Filter Statistics

《Adversarial Examples Detection in Deep Networks with Convolutional Filter Statistics 》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、AdversarialExamplesDetectioninDeepNetworkswithConvolutionalFilterStatisticsXinLi,FuxinLiSchoolofElectricalEngineeringandComputerScienceOregonStateUniversityurumican@gmail.com,lif@eecs.oregonstate.eduAbstractDeeplearninghasgreatlyimprovedvisualrecognitioninrecen

2、tyears.However,recentresearchhasshownthatthereexistmanyadversarialexamplesthatcannegativelyimpacttheperformanceofsuchanarchitecture.Thispaperfocusesondetectingthoseadversarialexamplesbyanalyz-ingwhethertheycomefromthesamedistributionasthenormalexamples.Insteado

3、fdirectlytrainingadeepneuralFigure1.Anoptimizationalgorithmcanfindtheadversarialex-networktodetectadversarials,amuchsimplerapproachisamplewhere,withalmostnegligibleperturbationstohumaneyes,proposedbasedonstatisticsonoutputsfromconvolutionalwillcompletelydistortt

4、hepredictionresultofadeepneuralnet-layers.Acascadeclassifierisdesignedtoefficientlydetectwork[26].Thisalgorithmisquiteuniversal,havingbeensuccess-adversarials.Furthermore,trainedfromoneparticularad-fullytestedonmanydifferentnetworksandtheusercandirecttheversarial

5、generatingmechanism,theresultingclassifiercannetworktooutputanycategorywithadversarialoptimization.successfullydetectadversarialsfromacompletelydifferentmechanismaswell.Afterdetectingadversarialexamples,andotherdevastatingeffectswouldbeunavoidable.weshowthatmany

6、ofthemcanberecoveredbysimplyper-formingasmallaveragefilterontheimage.Thosefind-Therefore,thereareamplereasonstobelievethatitisingsshouldprovokeustothinkmoreabouttheclassificationimportanttoidentifywhetheranexamplecomesfromanor-mechanismsindeepconvolutionalneuralne

7、tworks.maloranadversarialdistribution.Suchknowledgeifavail-ablewillhelpsignificantlytocontrolbehaviorsofrobotsemployingdeeplearning.Areliableprocedurecanpreventrobotsfrombehavinginundesirablemannersundesirable1.Introductionbecauseofthefalseperceptionsitmadeabout

8、theenviron-Recentadvancesindeeplearninghavegreatlyimprovedment.arXiv:1612.07767v1[cs.CV]22Dec2016thecapabilitytorecognizeimagesautomatically[13,24,8].Theunderstandingofwheth