国际会议演讲稿Word格式文档下载.docx
《国际会议演讲稿Word格式文档下载.docx》由会员分享,可在线阅读,更多相关《国际会议演讲稿Word格式文档下载.docx(6页珍藏版)》请在冰豆网上搜索。
P4
Partone,introduction
Firstly,Iwouldliketogiveyouabitofbackground.
Differingfromthetraditionalcryptographictechniquesbasedonsecretkeys,wecanmakeuseofwirelesschannelcharacteristicstoenhancephysicallayersecurity.
Cooperativecommunicationhasbeenwidelyrecognizedasaneffectivewaytocombatwirelessfadingandprovidediversitygainwhichisoneoftheresearchhotspots.
Machinelearningasanemergingtechnologyhasbeenwidelyappliedinimageprocessing,cancerprediction,stockanalysisandotherfields.Sowhynottryitinwirelesscommunication?
P5:
Next,Iwanttotalkalittlebitaboutpresentstudy
Recentstudiesondeeplearningforwirelesscommunicationsystemshaveproposedalternativeapproachestoenhancecertainpartsoftheconventionalcommunicationsystemsuchasmodulationrecognition、channelencodinganddecoding、channelestimationanddetectionandanautoencoderwhichcanreplacethetotalsystemwithanovelarchitecture
【modulationrecognition:
AnNNarchitectureformodulationrecognitionthatconsistsofa4-layerNNandtwotwo-layerNNs。
channelencodinganddecoding:
AplainDNNarchitectureforchanneldecodingtodecodekbits
messagesfromNbitsnoisycodewords。
channelestimationanddetection:
Adense-Netforsymbol-to-symboldetectioncanadoptlongshort-termmemory(LSTM)todetectanestimatedsymbol.
Autoencoder:
theautoencodercanrepresenttheentirecommunicationsystemandjointlyoptimizethetransmitterandreceiveroveranAWGNchannel.
】
P6
Sowhydidweconductthisresearch?
Well,wewanttoexploitthepotentialbenefitsofdeeplearninginenhancingphysicallayersecurityincooperative(
/kəʊ'
ɒpərətɪv/
wirelesscommunicationandreducethefeedbackoverheadinlimitedspectrumresoucebyourourproposedscheme.
P8
Nowletmemoveontoparttwo-systemmodel
Here,youcanseeafigurewhichisasystemmodel.
Thisisthesource;
thesearetherelaynodesandthisisthedestination,thisistheeavesdropper
Thewholeprocessofcooperativewirelesscommunicationcanbedividedintotwophases.
Inthefirstphase,thesourcebroadcaststhesignaltotheoptimalrelaywhichguaranteesperfectsecurity.AsshowninFig1,
representsafadingcoefficientofthechannelfromthesourcetotherelaynode(
.)
Inthesecondphase,theoptimalrelayforwardsascaledversionofitsreceivedsignaltothedestinationinthepresenceoftheeavesdropper,wheretheoptimalrelayisconsideredtoadoptamplify-and-forward(AF)relayscheme.
Inthisfigure,
representsafadingcoefficientofthechannelfromtherelay
tothedestination
representsafadingcoefficientofthechannelfromtherelay
totheeavesdropper.
P9:
Hereyoucanseesomefollowingexpressions.Iamnotgoingtowasteourprecioustimeonthelengthyderivation.Iwouldliketoinviteyoutodirectlytakealookattheequationinitsfinalform.
Thisistheoptimalindexoftheselectedrelaywiththeconventionalrelayselectionscheme.Amaongthisexpression
representstheachievablesecrecyrateofsystemmodelwhenthe
relayisselected.
P11
Nowletmemovetopartthree-----NN-basedRelaySelection
Hereyoucanseeafigurewhichshowsconventional3-layerneuralnetwork.Itconsistsofinputlayer,hiddenlayer1,hidden(/'
hɪdn/)layer2andoutputlayer.Neuralnetworkcanlearnfeaturesfromrawdataautomaticallyandadjustparameters(/pəˈræ
mɪtə(r)z/)flexibly(
/'
fleksəbli/)suchasweightsandbiases.
Incomplex(
kɒmpleks/)conditions(scenarios(/sɪ'
nɑːrɪəʊ/),)Neuralnetworkhaspromisingapplicationsinrelayselectionforseveralreasons.
First,thedeepnetworkhassuperior(/suːˈpɪərɪə/)learningabilitydespite(/dɪ'
spaɪt/)thecomplexchannelconditions.
Second,Neuralnetworkcanhandlelargedatasetsbecauseofdistributed(/dɪ'
strɪbjʊtɪd/)andparallel(/'
pæ
rəlel/)computing(/kəm'
pjuːtɪŋ/s,whichensurecomputation(/kɒmpjʊ'
teɪʃ(ə)n/)speedandprocessingcapacity(
/kə'
sɪtɪ/).
Third,variouslibrariesorframeworks,suchasTensorFlow,Theano,andCaffegiveitwideapplications
Inthispaper,theproblemoftherelayselectionismodeledasamulti(/'
mʌltɪ/,ao)-classificationproblem.Weadoptsimpleneuralnetwork(NN)toselecttheoptimalrelaytoguaranteesperfectsecrecyperformanceofrelaycooperativecommunicationsystem.(enhancephysicallayersecurity)
P12
Beforetrainingtheclassificationmodel,weneedtomakesomepreparationfordeeplearningtoacquireatrainingsetandatestingset.
First,weneedtoproducerealfeaturevectorforeachexampleaccordingtochannelstateinformation;
becausethechannelstateinformationmatricesiscomposedofcomplexnumbersbutfeaturevectorsaregenerallycomposedofrealnumbers.Soweneedtochangecomplexnumbersintorealnumberswithabsolute(/'
æ
bsəluːt/)valueoperation.Moreover,inordertoimprovetheclassificationperformance(precision),itisnecessarytonormalizethefeaturevectors.
Second,weneedtodesignkeyperformanceindicator(KPI).Inordertoeffectivelypreventtheeavesdropperfrominterceptinginformation,wechooseachievablesecrecyrateastheKPIofsystem.ThisKPIindicates(represents/shows)thedifferenceoftheachievableratefromthesourcetothedestinationandtheachievableratefromthesourcetotheeavesdropper.
Third,wecanmakelabelsforexamplesaccordingtoKPI.theindexoftherelaywhichobtainsthemaximum(
mæ
ksɪməm/)KPIisregardedastheclasslabeloftheexample.
P13
Classificationmodel
Thispicture(isabout)showsthewholeprocessofbuildingclassificationmodel.
Thewholeprocessofbuildingclassificationcanbedividedintotwophases,namelytrainingphaseandtestingphase.Inthefirstphase,weneedtochoosesuitablehyper(
haɪpə/)parameterstotrainneuralnetworkmodel.Inthesecondphase,wecanpredict(/prɪ'
dɪkt/)labelsofoptimalrelayaccordingtoinputdataandassessclassificationperformance.
P15
Nowletmemovetopartfour-----SimulationandResults(
/rɪˈzʌlts/)Analysis
Here,youcanseeafigurewhichshowstherelationshipbetweentheaveragetransmit(
/træ
nzˈmɪt/)powerofthesourceandtheachievablesecrecyratewithdifferentnumbersofrelays.
Inthisfigure,thebluelinerepresentstheconventionalrelayselectionschemeandtheredlinerepresentstheNN-basedscheme.
Inthisfigure,asthenumbersofrelayandtheaveragetransmitpowerofthesourceincreases,theachievablesecrecyrateincreasesaccordingly,.whichmeansincreasingthenumberofrelayscaneffectivelyimprovethesecretperformance.
Theredlinearealmost(
ɔːlməʊst/)closetotheblueline,【whchindicatesthatourproposedscheme(i.e.theNN-basedscheme)achievesalmostthesamesecrecyratesasthoseoftheconventionalschemeforallvaluesof
】whichvalidateseffectiveness(/ɪ'
fektɪvnɪs/)ofourproposedschemes.
P16
Thistableshowsthethenormalized(/ˈnɔrməˌlaɪzd/)meansquare(
/skweə/)error(NMSES)valuesofdirrerentrelaynodes.ThevalueofNMSEmeanstheperformancedifferencebetweentheconventionalschemeandourproposedscheme.ThevaluesofNMSEarebelow(/bɪ'
ləʊ/)negative('
negətɪv/)20(
),whichvalidateseffectivenessofourproposedschemeagain.
P17
Now,letmemovetothelastpart-----Conclusion
Okay,nowwearegoingtotakealookatthelastpart-Conclusion.
P18
Wehavegotthefollowingconclusions.
First,Incomplex(conditions)scenarios,Neuralnetworkhaspromisingapplicationsinrelayselectionforsuperiorlearningability,computationspeedandprocessingcapacity.
Second,Comparedwiththeconventionalrelayselectionscheme,ourproposedschemeachievesalmostthesamesecrecyperformance.
Andlast,Ourproposedschemehasanadvantage(/əd'
vɑːntɪdʒ/)ofrelativelysmallfeedbackoverhead,indicatingthatproposedschemecanbeappliedtotheconditions(scenarios)wherethefeedbackislimited.
(Iftheconventionalschemeneedsfeedbackof
complexnumbers,NN-basedschemewillonlyneedfeedbackof
realnumbers.Therefore,thefeedbackoverheadofourproposedschemeishalf(/hɑːf/
)ofthatoftheconventionalscheme,)
Q&
A
1、计算复杂度
Computationalcomplexity
Thebiggestdrawbackisthehighlyselectioncomplexitieswithasmallnumberofrelaynodes.
Ifnumberofrelaynodeisbig,itwillhaveaadvantage.Thisneedourfurtherresearch.
Q:
TheexperimentshowsthatsecrecyrateisalmostthesameastraditionalmethodandwhatisthepromotionofusingNNtorelayselection.(whatismeaningofintroducingNNtorelayselection)
A:
Thatourproposedscheme(i.e.theNN-basedscheme)achievesalmostthesameachievablesecrecyrateasthatoftheconventionalschemeindicatesthatourproposedschemeiseffectiveanditcanselectoptimalrelaynodewhichobtainsmaximumachievablesecrecyrate.
Onereason(thefirstreason)isthatAdoptingNNforrelayselectionisanovelidea.
Anotherreasonisthatthespectrumresourceisrelativelimitedandourproposedschemehassmallfeedbackoverhead.
what’sthemeaningof“perfectsecrecyperformance”?
What’sthemeaningof“Comparedtotheconv