poster print size: improving credit card fraud deteccon...

1
Improving Credit Card Fraud DetecCon using CNN/GAN Rajesh Sabari, [email protected] Rajesh Sabari [email protected] Contact 1.S. Maes, K. Tuyls, B. Vanschoenwinkel, and B. Manderick, “Credit Card Fraud DetecAon Using Bayesian and Neural Networks,” 2. K. Fu, D. Cheng, Y. Tu, and L. Zhang, “Credit Card Fraud DetecAon Using ConvoluAonal Neural Networks,” References Using GAN for 5000 rounds piqng the generator network against the discriminator network, making use of the cross-entropy loss from the discriminator to train the networks for improving classificaAon effecAveness in Credit Card Fraud DetecAon. The augmented image is passed through 1x29 convoluAonal layer followed by a fully connected dense layers to finally have a SoGmax predictor MoCvaCon/Summary Dataset The base line models were LogisAc Regression, Random Forest and GaussianNB. 3 DL models used are 2 layer MLP, GAN, two 1D Conv layer , max pooling, a fully connected (300 neurons) and soGmax classifier (2 classes). prediction). Models CNN (best performance): 0.860230099502 RandomForestClassifier:0.846437 (baseline) MLPClassifier (2 Layer with drop out): 0.8085106382978723 MLPClassifier (1 Layer): 0.707243346007604 Discussion CNN with enhanced fraud dataset through GAN yielded high training and test accuracy on our training and test sets. Adding drop out certainly yielded lower variance and addiAonal layers in MLP enhanced performance. The poor generalizaAon is a result of our dataset not being diverse. By the end of 5000 training iteraAons the generated fraud imaged parern started to mimic actual fraud Conclusions Model Results Datasets used for training in this project are from Kaggle with 31 features including the Ame and amount of transacAon as well as a label whether that transacAon was fraudulent or not. The variables are PCA transformed. 99.83% of transacAons in this dataset were not fradulent while only 0.17% were fradulent

Upload: others

Post on 15-Oct-2020

5 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Poster Print Size: Improving Credit Card Fraud DetecCon ...cs230.stanford.edu/projects_winter_2020/posters/32635168.pdf · Poster Print Size: This poster template is 24” high by

PosterPrintSize:Thispostertemplateis24”highby36”wide.Itcanbeusedtoprintanyposterwitha2:3aspectraAoincluding36x54and48x72.

Placeholders:ThevariouselementsincludedinthisposterareonesweoGenseeinmedical,research,andscienAficposters.Feelfreetoedit,move,add,anddeleteitems,orchangethelayouttosuityourneeds.Alwayscheckwithyourconferenceorganizerforspecificrequirements.

ImageQuality:YoucanplacedigitalphotosorlogoartinyourposterfilebyselecAngtheInsert,Picturecommand,orbyusingstandardcopy&paste.Forbestresults,allgraphicelementsshouldbeatleast150-200pixelsperinchintheirfinalprintedsize.Forinstance,a1600x1200pixelphotowillusuallylookfineupto8“-10”wideonyourprintedposter.Topreviewtheprintqualityofimages,selectamagnificaAonof100%whenpreviewingyourposter.Thiswillgiveyouagoodideaofwhatitwilllooklikeinprint.Ifyouarelayingoutalargeposterandusinghalf-scaledimensions,besuretopreviewyourgraphicsat200%toseethemattheirfinalprintedsize.

Pleasenotethatgraphicsfromwebsites(suchasthelogoonyourhospital'soruniversity'shomepage)willonlybe72dpiandnotsuitableforprinAng.

[Thissidebarareadoesnotprint.]

ChangeColorTheme:Thistemplateisdesignedtousethebuilt-incolorthemesinthenewerversionsofPowerPoint.Tochangethecolortheme,selecttheDesigntab,thenselecttheColorsdrop-downlist.

Thedefaultcolorthemeforthistemplateis“Office”,soyoucanalwaysreturntothataGertryingsomeofthealternaAves.

PrinAngYourPoster:Onceyourposterfileisready,visitwww.genigraphics.comtoorderahigh-quality,affordableposterprint.EveryorderreceivesafreedesignreviewandwecandeliverasfastasnextbusinessdaywithintheUSandCanada.Genigraphics®hasbeenproducingoutputfromPowerPoint®longerthananyoneintheindustry;daAngbacktowhenwehelpedMicrosoG®designthePowerPoint®soGware.

USandCanada:1-800-790-4001Email:[email protected]

[Thissidebarareadoesnotprint.]

ImprovingCreditCardFraudDetecConusingCNN/GAN

RajeshSabari,[email protected]

[email protected]

Contact1. S.Maes,K.Tuyls,B.Vanschoenwinkel,andB.Manderick,“CreditCardFraudDetecAonUsingBayesianandNeuralNetworks,”

2. K.Fu,D.Cheng,Y.Tu,andL.Zhang,“CreditCardFraudDetecAonUsingConvoluAonalNeuralNetworks,”

References

UsingGANfor5000roundspiqngthegeneratornetworkagainstthediscriminatornetwork,makinguseofthecross-entropylossfromthediscriminatortotrainthenetworksforimprovingclassificaAoneffecAvenessinCreditCardFraudDetecAon.Theaugmentedimageispassedthrough1x29convoluAonallayerfollowedbyafullyconnecteddenselayerstofinallyhaveaSoGmaxpredictor

MoCvaCon/Summary

Dataset

ThebaselinemodelswereLogisAcRegression,RandomForestandGaussianNB.3DLmodelsusedare2layerMLP,GAN,two1DConvlayer,maxpooling,afullyconnected(300neurons)andsoGmaxclassifier(2classes). prediction).

Models

CNN(bestperformance):0.860230099502RandomForestClassifier:0.846437(baseline)

MLPClassifier(2Layerwithdropout):0.8085106382978723MLPClassifier(1Layer):0.707243346007604

Discussion

CNNwithenhancedfrauddatasetthroughGANyieldedhightrainingandtestaccuracyonourtrainingandtestsets.AddingdropoutcertainlyyieldedlowervarianceandaddiAonallayersinMLPenhancedperformance.ThepoorgeneralizaAonisaresultofourdatasetnotbeingdiverse.Bytheendof5000trainingiteraAonsthegeneratedfraudimagedparernstartedtomimicactualfraud

Conclusions

ModelResults

DatasetsusedfortraininginthisprojectarefromKagglewith31featuresincludingtheAmeandamountoftransacAonaswellasalabelwhetherthattransacAonwasfraudulentornot.ThevariablesarePCAtransformed.99.83%oftransacAonsinthisdatasetwerenotfradulentwhileonly0.17%werefradulent