第47卷第4期电网技术Vol.47No.42023年4月PowerSystemTechnologyApr.2023文章编号：1000-3673（2023）04-1531-09中图分类号：TM721文献标志码：A学科代码：470·40低成本对抗性隐蔽虚假数据注入攻击及其检测方法黄冬梅1，丁仲辉2，胡安铎1，王晓亮3，时帅2（1．上海电力大学电子与信息工程学院，上海市浦东新区201306；2．上海电力大学电气工程学院，上海市杨浦区200090；3．国家海洋局东海信息中心，上海市浦东新区200136）Low-costAdversarialStealthyFalseDataInjectionAttackandDetectionMethodHUANGDongmei1,DINGZhonghui2,HUAnduo1,WANGXiaoliang3,SHIShuai2(1.CollegeofElectronicsandInformationEngineering,ShanghaiUniversityofElectricPower,PudongNewArea,Shanghai201306,China;2.CollegeofElectricalEngineering,ShanghaiUniversityofElectricPower,YangpuDistrict,Shanghai200090,China;3.EastSeaInformationCenter,StateOceanicAdministration,PudongNewArea,Shanghai200136,China)ABSTRACT:Powergridswithdeepcouplingsinthephysicalandinformativeaspectsfacethethreatoffalsedatainjectionattacks,whilethedeeplearningtechniquebecomesanimportantmethodfordetectingthefalsedatainjectionattacks.Toaddresstheproblemthatthedeepneuralnetworksarevulnerabletoadversarialattacks,alow-costadversarialtargetoptimalconcealedfalsedatainjectionattackstrategiesandthecorrespondingdetectionmethodsareproposedinthispaper.Theoptimalcombinationofmeasurementsandattackvaluearesolvedbyatwo-stageoptimizationtoobtaintheoptimalattackstrategywiththelowestattackcostfortheexpectedattacktarget.Theoptimalattackvalueisaddedwiththeadversarialperturbationbythewhite-boxattacksothatthedeeplearningmodelincorrectlyreportsitasanormalsample.Fromthedetectionperspective,allinitialattacksamplesinthehistoricaldatabaseareaddedwiththeadversarialperturbationandaremarkedastheattacksamplessothattheyarethenaddedtothetrainingsettotrainthemodelandimprovethedetectionperformance.ExperimentsareconductedontheIEEE14-nodesystemandtheIEEE118-nodesystemrespectivelytoverifytheeffectivenessoftheproposedattackanddetectionmethods.KEYWORDS:deeplearning;falsedatainjectionattack;adversarialattack;white...
