摘要
Facede-identificationhasbecomeincreasinglyimportantastheimagesourcesareexplosivelygrowingandeasilyaccessible.Theadvanceofnewfacerecognitiontechniquesalsoarisespeople'sconcernregardingtheprivacyleakage.Themainstreampipelinesoffacede-identificationaremostlybasedonthek-sameframework,whichbearscritiquesofloweffectivenessandpoorvisualquality.Inthispaper,weproposeanewframeworkcalledPrivacy-Protective-GAN(PP-GAN)thatadaptsGAN(generativeadversarialnetwork)withnovelverificatorandregulatormodulesspeciallydesignedforthefacede-identificationproblemtoensuregeneratingde-identifiedoutputwithretainedstructuresimilarityaccordingtoasingleinput.Weevaluatetheproposedapproachintermsofprivacyprotection,utilitypreservation,andstructuresimilarity.Ourapproachnotonlyoutperformsexistingfacede-identificationtechniquesbutalsoprovidesapracticalframeworkofadaptingGANwithpriorsofdomainknowledge.
出版日期
2019年01月11日(中国Betway体育网页登陆平台首次上网日期,不代表论文的发表时间)