自我介紹Thankyou,Mr./Ms.Chair./professorMynameissangqian.Iamveryhonoredtobeheretodooralpresentation.IamaMasterstudentfromHohaiUniversityandIamcurrentlydoingsomeresearchonphysicallayersecurity.Today,Iwouldliketosharewithyousomeofmyresearchonrelayselectionincooperativecommunication.(external/ek?st?rn?l;?k?st?rn?l/)內(nèi)容安排：Mypresentationincludesthesefiveparts.First,somebackgroundinformationaboutthisresearch;Second,systemmodelwehavedone;Third,NN-basedrelayselectionschemewehaveproposedForth,SimulationandresultsanalysisAndlast,someconclusionswehavegotP4Partone,introductionFirstly,Iwouldliketogiveyouabitofbackground.Differingfromthetraditionalcryptographictechniquesbasedonsecretkeys,wecanmakeuseofwirelesschannelcharacteristicstoenhancephysicallayersecurity.Cooperativecommunicationhasbeenwidelyrecognizedasaneffectivewaytocombatwirelessfadingandprovidediversitygainwhichisoneoftheresearchhotspots.Machinelearningasanemergingtechnologyhasbeenwidelyappliedinimageprocessing,cancerprediction,stockanalysisandotherfields.Sowhynottryitinwirelesscommunication?P5:Next,IwanttotalkalittlebitaboutpresentstudyRecentstudiesondeeplearningforwirelesscommunicationsystemshaveproposedalternativeapproachestoenhancecertainpartsoftheconventionalcommunicationsystemsuchasmodulationrecognition、channelencodinganddecoding、channelestimationanddetectionandanautoencoderwhichcanreplacethetotalsystemwithanovelarchitecture【modulationrecognition:AnNNarchitectureformodulationrecognitionthatconsistsofa4-layerNNandtwotwo-layerNNs。channelencodinganddecoding：AplainDNNarchitectureforchanneldecodingtodecodekbitsmessagesfromNbitsnoisycodewords。channelestimationanddetection：Adense-Netforsymbol-to-symboldetectioncanadoptlongshort-termmemory(LSTM)todetectanestimatedsymbol.Autoencoder：theautoencodercanrepresenttheentirecommunicationsystemandjointlyoptimizethetransmitterandreceiveroveranAWGNchannel.】P6Sowhydidweconductthisresearch?Well,wewanttoexploitthepotentialbenefitsofdeeplearninginenhancingphysicallayersecurityincooperative(

/k??'?p?r?t?v/

wirelesscommunicationandreducethefeedbackoverheadinlimitedspectrumresoucebyourourproposedscheme.P8P9:Hereyoucanseesomefollowingexpressions.Iamnotgoingtowasteourprecioustimeonthelengthyderivation.Iwouldliketoinviteyoutodirectlytakealookattheequationinitsfinalform.Thisistheoptimalindexoftheselectedrelaywiththeconventionalrelayselectionscheme.Amaongthisexpressionrepresentstheachievablesecrecyrateofsystemmodelwhentherelayisselected.P11Hereyoucanseeafigurewhichshowsconventional3-layerneuralnetwork.Itconsistsofinputlayer,hiddenlayer1,hidden(/'h?dn/)layer2andoutputlayer.Neuralnetworkcanlearnfeaturesfromrawdataautomaticallyandadjustparameters（/p??r?m?t?(r)z/）flexibly(

/'fleks?bli/)suchasweightsandbiases.Incomplex(

/'k?mpleks/)conditions(scenarios(/s?'nɑ?r???/),)Neuralnetworkhaspromisingapplicationsinrelayselectionforseveralreasons.First,thedeepnetworkhassuperior(/su??p??r??/)learningabilitydespite(/d?'spa?t/)thecomplexchannelconditions.Second,Neuralnetworkcanhandlelargedatasetsbecauseofdistributed(/d?'str?bj?t?d/)andparallel(/'p?r?lel/)computing(/k?m'pju?t??/s,whichensurecomputation(/k?mpj?'te??(?)n/)speedandprocessingcapacity(

/k?'p?s?t?/).Third,variouslibrariesorframeworks,suchasTensorFlow,Theano,andCaffegiveitwideapplicationsInthispaper,theproblemoftherelayselectionismodeledasamulti(/'m?lt?/,ao)-classificationproblem.Weadoptsimpleneuralnetwork(NN)toselecttheoptimalrelaytoguaranteesperfectsecrecyperformanceofrelaycooperativecommunicationsystem.(enhancephysicallayersecurity)P12Beforetrainingtheclassificationmodel,weneedtomakesomepreparationfordeeplearningtoacquireatrainingsetandatestingset.First,weneedtoproducerealfeaturevectorforeachexampleaccordingtochannelstateinformation;becausethechannelstateinformationmatricesiscomposedofcomplexnumbersbutfeaturevectorsaregenerallycomposedofrealnumbers.Soweneedtochangecomplexnumbersintorealnumberswithabsolute(/'?bs?lu?t/)valueoperation.Moreover,inordertoimprovetheclassificationperformance(precision),itisnecessarytonormalizethefeaturevectors.Third,wecanmakelabelsforexamplesaccordingtoKPI.theindexoftherelaywhichobtainsthemaximum(

/'m?ks?m?m/)KPIisregardedastheclasslabeloftheexample.ClassificationmodelThispicture(isabout)showsthewholeprocessofbuildingclassificationmodel.Thewholeprocessofbuildingclassificationcanbedividedintotwophases,namelytrainingphaseandtestingphase.Inthefirstphase,weneedtochoosesuitablehyper(

/'ha?p?/)parameterstotrainneuralnetworkmodel.Inthesecondphase,wecanpredict(/pr?'d?kt/)labelsofoptimalrelayaccordingtoinputdataandassessclassificationperformance.P15Nowletmemovetopartfour-----SimulationandResults(

/r??z?lts/)AnalysisHere,youcanseeafigurewhichshowstherelationshipbetweentheaveragetransmit(

/tr?nz?m?t/)powerofthesourceandtheachievablesecrecyratewithdifferentnumbersofrelays.Theredlinearealmost(

/'??lm??st/)closetotheblueline,【whchindicatesthatourproposedscheme(i.e.theNN-basedscheme)achievesalmostthesamesecrecyratesasthoseoftheconventionalschemeforallvaluesof,】whichvalidateseffectiveness(/?'fekt?vn?s/)ofourproposedschemes.Thistableshowsthethenormalized(/?n?rm??la?zd/)meansquare(

/skwe?/)error(NMSES)valuesofdirrerentrelaynodes.ThevalueofNMSEmeanstheperformancedifferencebetweentheconventionalschemeandourproposedscheme.ThevaluesofNMSEarebelow(/b?'l??/)negative('neg?t?v/)20(),whichvalidateseffectivenessofourproposedschemeagain.P18Wehavegotthefollowingconclusions.First,Incomplex(conditions)scenarios,Neuralnetworkhaspromisingapplicationsinrelayselectionforsuperiorlearningability,computationspeedandprocessingcapacity.Second,Comparedwiththeconventionalrelayselectionscheme,ourproposedschemeachievesalmostthesamesecrecyperformance.Andlast,Ourproposedschemehasanadvantage(/?d'vɑ?nt?d?/)ofrelativelysmallfeedbackoverhead,indicatingthatproposedschemecanbeappliedtotheconditions(scenarios）wherethefeedbackislimited.(Iftheconventionalschemeneedsfeedbackofcomplexnumbers,NN-basedschemewillonlyneedfeedbackofrealnumbers.Therefore,thefeedbackoverheadofourproposedschemeishalf(/hɑ?f/

)ofthatoftheconventionalscheme,)Q&A計(jì)算復(fù)雜度ComputationalcomplexityThebiggestdrawbackisthehighlyselectioncomplexitieswithasmallnumberofrelaynodes.Ifnumberofrelaynodeisbig,itwillhaveaadvantage.Thisneedourfurtherresearch.Q:TheexperimentshowsthatsecrecyrateisalmostthesameastraditionalmethodandwhatisthepromotionofusingNNtorelayselection.(whatismeaningofintroducingNNtorelayselection)A:Thatourproposedscheme(i.e.theNN-basedscheme)achievesalmostthesameachievablesecrecyrateasthatoftheconventionalschemeindicatesthatourproposedschemeiseffectiveanditcanselectoptimalrelaynodewhichobtainsmaximumachievablesecrecyrate.Onereason(thefirstreason)isthatAdoptingNNforrelayselectionisanovelidea.Anotherreasonisthatthespectrumresourceisrelativelimitedandourproposedschemehassmallfeedbackoverhead.Q:what’sthemeaningof“perfectsecrecyperformance”?What’sthemeaningof“Comparedtotheconventionalrelayselectionscheme”?A:“perfectsecrecyperformance”meanstheachievablesecrecyrateisthebiggestonewhichcanenhancephysicallayersecurity.Infact,theconventionalrelayselectionschemeistheexhaustivesearch.Theindexofrelayselectionwiththisschemeisthebestone.Q:“Itisobviousthatthefeedbackoverheadofproposedschemeishalfofthatoftheconventionalscheme”A:well,Let'smakeanassumption.Iftheconventionalschemeneedsfeedbackofcomplexnumbers,N