返回
2016-Adversarial Diversity and Hard Positive Generation

2016-Adversarial Diversity and Hard Positive Generation

ID：39713917

大小：1.38 MB

页数：8页

时间：2019-07-09

2016-Adversarial Diversity and Hard Positive Generation_第1页
2016-Adversarial Diversity and Hard Positive Generation_第2页
2016-Adversarial Diversity and Hard Positive Generation_第3页
2016-Adversarial Diversity and Hard Positive Generation_第4页
2016-Adversarial Diversity and Hard Positive Generation_第5页
资源描述：

《2016-Adversarial Diversity and Hard Positive Generation》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、2016IEEEConferenceonComputerVisionandPatternRecognitionWorkshopsAdversarialDiversityandHardPositiveGenerationAndrasRozsa,EthanM.Rudd,andTerranceE.Boult∗UniversityofColoradoatColoradoSpringsVisionandSecurityTechnology(VAST)Lab{arozsa,erudd,tboult}@vast.uccs.eduAbstract  State-of-t

2、he-artdeepneuralnetworkssufferfromafun-damentalproblem–theymisclassifyadversarialexamplesformedbyapplyingsmallperturbationstoinputs.Inthispaper,wepresentanewpsychometricperceptualadversar-ialsimilarityscore(PASS)measureforquantifyingadver-sarialimages,introducethenotionofha

3、rdpositivegenera-tion,anduseadiversesetofadversarialperturbations–notjusttheclosestones–fordataaugmentation.Weintroduceanovelhot/coldapproachforadversarialexamplegener-ation,whichprovidesmultiplepossibleadversarialpertur-bationsforeverysingleimage.TheperturbationsgeneratedFigure1:ADVERSARIALSAND

4、HARDPOSITIVES.Thisbyournovelapproachoftencorrespondtosemanticallypaperdemonstrateshowtogenerateamuchmorediversesetmeaningfulimagestructures,andallowgreaterflexibilityofadversarialexamplesthantheexistingL-BFGS[21]orfasttoscaleperturbation-amplitudes,whichyieldsanincreasedgradientsign(FGS)[8]method

5、s.Viaarangeofperturbationdiversityofadversarialimages.Wepresentadversarialamplitudesalongthelearntadversarialdirections–notjusttheclosestadversarialsample–wecangeneratehardpositivestoimagesonseveralnetworktopologiesanddatasets,includ-fine-tunetheclassdefinitions,therebyextendingpreviouslyover-ingL

6、eNetontheMNISTdataset,andGoogLeNetandfitdecisionboundariestoimprovebothaccuracyandrobustness.ResidualNetontheImageNetdataset.Finally,wedemon-Theextendeddecisionboundariesarerepresentedbydashedlines.strateonLeNetandGoogLeNetthatfine-tuningwithaThissimplifiedschematicusesshapestodepictdifferenttypeso

7、fdiversesetofhardpositivesimprovestherobustnessofthesehardpositiveexamples.Innercolorsdepicttheoriginalclass,networkscomparedtotrainingwithpriormethodsofgen-whileoutercolorsdepicttheclassificationbythebasenetwork.eratingadver

当前文档最多预览五页，下载文档查看全文

此文档下载收益归作者所有

当前文档最多预览五页，下载文档查看全文
温馨提示：
1. 部分包含数学公式或PPT动画的文件，查看预览时可能会显示错乱或异常，文件下载后无此问题，请放心下载。
2. 本文档由用户上传，版权归属用户，天天文库负责整理代发布。如果您对本文档版权有争议请及时联系客服。
3. 下载前请仔细阅读文档内容，确认文档内容符合您的需求后进行下载，若出现内容与标题不符可向本站投诉处理。
4. 下载文档时可能由于网络波动等原因无法下载或下载错误，付费完成后未能成功下载的用户请联系客服处理。