資源描述：
《Adversarial Examples Detection in Deep Networks with Convolutional Filter Statistics 》由會員上傳分享，免費(fèi)在線閱讀，更多相關(guān)內(nèi)容在學(xué)術(shù)論文-天天文庫。
1、AdversarialExamplesDetectioninDeepNetworkswithConvolutionalFilterStatisticsXinLi,FuxinLiSchoolofElectricalEngineeringandComputerScienceOregonStateUniversityurumican@gmail.com,lif@eecs.oregonstate.eduAbstractDeeplearninghasgreatlyimprovedvisualrecognitioninrecen
2、tyears.However,recentresearchhasshownthatthereexistmanyadversarialexamplesthatcannegativelyimpacttheperformanceofsuchanarchitecture.Thispaperfocusesondetectingthoseadversarialexamplesbyanalyz-ingwhethertheycomefromthesamedistributionasthenormalexamples.Insteado
3、fdirectlytrainingadeepneuralFigure1.Anoptimizationalgorithmcan?ndtheadversarialex-networktodetectadversarials,amuchsimplerapproachisamplewhere,withalmostnegligibleperturbationstohumaneyes,proposedbasedonstatisticsonoutputsfromconvolutionalwillcompletelydistortt
4、hepredictionresultofadeepneuralnet-layers.Acascadeclassi?erisdesignedtoef?cientlydetectwork[26].Thisalgorithmisquiteuniversal,havingbeensuccess-adversarials.Furthermore,trainedfromoneparticularad-fullytestedonmanydifferentnetworksandtheusercandirecttheversarial
5、generatingmechanism,theresultingclassi?ercannetworktooutputanycategorywithadversarialoptimization.successfullydetectadversarialsfromacompletelydifferentmechanismaswell.Afterdetectingadversarialexamples,andotherdevastatingeffectswouldbeunavoidable.weshowthatmany
6、ofthemcanberecoveredbysimplyper-formingasmallaverage?lterontheimage.Those?nd-Therefore,thereareamplereasonstobelievethatitisingsshouldprovokeustothinkmoreabouttheclassi?cationimportanttoidentifywhetheranexamplecomesfromanor-mechanismsindeepconvolutionalneuralne
7、tworks.maloranadversarialdistribution.Suchknowledgeifavail-ablewillhelpsigni?cantlytocontrolbehaviorsofrobotsemployingdeeplearning.Areliableprocedurecanpreventrobotsfrombehavinginundesirablemannersundesirable1.Introductionbecauseofthefalseperceptionsitmadeabout
8、theenviron-Recentadvancesindeeplearninghavegreatlyimprovedment.arXiv:1612.07767v1[cs.CV]22Dec2016thecapabilitytorecognizeimagesautomatically[13,24,8].Theunderstandingofwheth
