- 1、本文档共54页,可阅读全部内容。
- 2、原创力文档(book118)网站文档一经付费(服务费),不意味着购买了该文档的版权,仅供个人/单位学习、研究之用,不得用于商业用途,未经授权,严禁复制、发行、汇编、翻译或者网络传播等,侵权必究。
- 3、本站所有内容均由合作方或网友上传,本站不对文档的完整性、权威性及其观点立场正确性做任何保证或承诺!文档内容仅供研究参考,付费前请自行鉴别。如您付费,意味着您自己接受本站规则且自行承担风险,本站不退款、不进行额外附加服务;查看《如何避免下载的几个坑》。如果您已付费下载过本站文档,您可以点击 这里二次下载。
- 4、如文档侵犯商业秘密、侵犯著作权、侵犯人身权等,请点击“版权申诉”(推荐),也可以打举报电话:400-050-0827(电话支持时间:9:00-18:30)。
AdversarialExampleDetection姜育刚,马兴军,吴祖煊
Recap:week31.AdversarialExamples2.AdversarialAttacks3.AdversarialVulnerabilityUnderstanding
In-classAdversarialAttackCompetitionhttps://codalab.lisn.upsaclay.fr/competitions/15669?secret_key=77cb8986-d5bd-4009-82f0-7dde2e819ff8
In-classAdversarialAttackCompetition
In-classAdversarialAttackCompetitionAdversarialattackcompetition(accountfor30%)必须使用学校邮箱注册比赛(否则无成绩)比赛时间:Phase1:10月1号–10月28号Phase2:评估阶段,学生不参与没卡的同学可以使用GoogleColab:/按排名算分:第一名30分最后一名15分
AdversarialExampleDetection(AED)Abinaryclassificationproblem:clean(y=0)oradv(y=1)?Ananomalydetectionproblem:benign(y=0)orabnormal(y=1)?
PrinciplesforAEDAllbinaryclassificationmethodscanbeappliedforAED
PrinciplesforAEDAllanomalydetectionmethodscanbeappliedforAED
PrinciplesforAEDUseasmuchinformationasyoucanInputstatisticsManualfeaturesTrainingdataAttentionmapTransformationMixupDenoising…ActivationsDeepfeaturesProbabilitiesLogitsGradientsLosslandscapeUncertainty…
PrinciplesforAEDLeverageuniquecharacteristicsofadversarialexamplesTwinsStrangersExtremelyclosetothecleansampleFarawayinprediction
PrinciplesforAEDBuilddetectorsbasedonexistingunderstandingsHighdimensionalpocketsLocallinearityTiltingboundary
PrinciplesforAEDIt’sisstillfeatureengineering!
ChallengesinAEDThediversityofadversarialexamplesusedfortrainingthedetectorsdeterminethedetectionperformanceDetectorsarealsomachinelearningmodels:theyarealsovulnerabletoadversarialattacksThedetectorsneedtodetectbothexistingandunknownattacksThedetectorsneedtoberobusttoadaptiveattacks
ExistingMethodsSecondaryClassificationMethods(二级分类法)PrincipleComponentAnalysis(主成分分析法,PCA)DistributionDetectionMethods(分布检测法)PredictionInconsistency(预测不一致性)ReconstructionInconsistency(重建不一致性)TrappingBasedDetection(诱捕检测法)
ExistingMethodsSecondaryClassificationMethods(二级分类法)PrincipleComp
文档评论(0)