a theoretical framework for robustness of ......a theoretical framework for robustness of (deep)...
TRANSCRIPT
ATHEORETICALFRAMEWORKFORROBUSTNESSOF(DEEP)CLASSIFIERSUNDERADVERSARIALEXAMPLES
BeilunWang,JiGaoandYanjun QiDepartmentofComputerScience,UniversityofVirginia
ProblemSetting:
DefineAdversarialExamples:
TowardsPrincipledSolutions(forDNNs):
OurtheoremssuggestalistofpossiblesolutionsthatmayimprovetherobustnessofDNNclassifiersagainstadversarialsamples.Optionsinclude,like(1)learningabetter12 ;(2)modifyingunnecessaryfeatures(SeePosterDeepMask-TuesdayMorningW18).
• For(1),thealternativemethodforhardeningtheDNNmodelsisminimizingsomelossfunctions345(7, 7′)sothatwhen:.(;. 7 , ;.(7′)) < =(approximatedby(>, ∥⋅∥)),thisloss345(7, 7′)issmall.Atableofcomparingexistinghardeningsolutionsusingthismethodisshownasfollowing:
ExperimentEvaluation
Define(AB, C)-Strong-robustness:
WhyDNNmodelisnotstrong-robust.
Whyaclassifierisvulnerabletoadversarialsamples.
SufficientConditionforStrong-robustness:
Strong-robustness forD.
ExperimentalEvaluation:
TowardsPrincipledUnderstanding