A Neural Network Based Model for Detecting Irregularities in e ...

Afr J Comp & ICTVol 4. No. 3. Issue 2Adeyiga et al – A Neural Network Based Model for Detecting Irregularities in e-Banking TransactionsA Neural Network Based Model for DetectingIrregularities in e-Banking Transactions1 J.A Adeyiga, J.O. Ezike, A. Omotosho & W. AmakulorDept. of Computer ScienceBells University of TechnologyOta, Nigeria1 Corresponding Authorjadeyiga@yahoo.comABSTRACTWith the migration of most banking operations to the online platform, predicting bank frauds andsecurity failures has gained increased importance in preventing and reducing the negative effects suchbreaches will have on the economic system as a whole. Information system security views this challengeas a prediction problem that attempts to detect irregular transactions in the banking sector operationsscenario. This study applies neural network techniques to the bank fraud prediction problem. UsingNigerian banks as a point of reference, we design a Neural Network-Based Model that employ multilayeredFeed Forward Artificial Neural Network on database system for collecting training data for the ArtificialNeural Network. Future work will test the Intelligence of the system on data extracted from statementsof accounts from different banks in Nigeria.Keywords: Artificial neural network, transactions, bank fraud, financial institutions & cyber security1. INTRODUCTIONIt is incontrovertible that the banking system is theengine of growth in any economy, given itsfunction of financial intermediation. Through thisfunction, banks facilitate capital formation andpromote economic growth. However, bank’s abilityto render economic growth and developmentdepends on the health, soundness and stability ofthe system [2]. It is, therefore, not surprising thatthe banking industry is one of the most regulatedsectors in any economy. It is against thisbackground that the Central Bank of Nigeriaoutlined as part of the first phase of its bankingsector reforms, the assurance of a diversified,strong and reliable banking industry [15].African Journal of Computing & ICT Reference Format:J.A Adeyiga, J.O.J Ezike, A. Omotosho & W. Amakulor (2011). ANeural Network Based Model for Detecting Irregularities in e-Banking Transactions. Afr J. of Comp & ICTs. Vol 4, No. 3.Issue 2 pp 7-14© African Journal of Computing & ICT December, 2011- ISSN 2006-1781The primary objective of the reforms is toguarantee an efficient and sound financial system.The reforms are designed to enable the bankingsystem develop the required resilience to supportthe economic development of the nation byefficiently performing its functions as the fulcrumof financial intermediation [11]. The objective is toensure the safety of depositors’ money, positionbanks to play active developmental roles in theNigerian economy, and become major players inthe sub-regional, regional and global financialmarkets.As more financial institutions in Nigeria movestowards information technology driven services,security issues need to be addressed before banksand customers can confidently take advantage ofthese platforms. Banks are afraid of losing theircash to fraudsters who can manipulate their systemand get undue advantage, while customers are stillnot convinced that banking is totally secured.There is need for customers’ transactions to bemonitored so as to notice and alert the bankofficials of irregular and suspicious transactions oncustomer accounts.7

Afr J Comp & ICTVol 4. No. 3. Issue 2Adeyiga et al – A Neural Network Based Model for Detecting Irregularities in e-Banking TransactionsCurrently, transactions are usually manuallymonitored by bank personnel who look through thecustomers’ statement of accounts when unusualtransactions are noticed. The process is often verytedious and inefficient mainly because of thenumber of transactions and customer base.With the advent of computers and Informationsystems there is a possibility of an automatedapproach to the analysis of the customers’statement of account for detecting irregularities inbanking activities. However, considering the fastpace at which mainstream business rules changes,the definition of fraudulent transaction changesrapidly thereby making the design anddevelopment of such a system a rather complexprocess [1]. A solution paradigm is to exploreautomated approaches to irregularity detectionusing algorithmic approach and artificial IntelligentSystem.2. RELATED WORKSIn [7][12] an approach to fraud detection that isbased on tracking calling behaviour on an accountover time and scoring calls according to the extentthat they deviate from patterns that resemblefraud are described. Account summaries arecompared to threshold each period and an accountwhose summary exceeds a threshold can be queuedto be analyzed for fraud. Thresholding has severaldisadvantages; it may vary with time of day, typeof account and types of call to be sensitive to fraudinvestigation without setting off too many falsealarms for legitimate traffic [12].Fawcett and Provost [7] developed an innovativemethod for choosing account-specific thresholdrather than universals threshold that apply to allaccounts or all accounts in a segment. In theexperiment, fraud detection is based on trackingaccount behaviour. Fraud detection was eventdriven and not time driven, so that fraud can bedetected as it is happening. Second, frauddetection must be able to learn the calling patternon an account and adapt to legitimate changes incalling behaviour. Lastly, fraud detection must beself-initializing so that it can be applied to newaccounts that do not have enough data for training.The approach adopted probability distributionfunctions to track legitimate calling behaviour.In [4] it was stated that that many financialinstitutions see the value of ANNs as a supportingmechanism for financial analysts and are activelyinvesting in this arena. The models describedprovide the needed knowledge to choose the typeof neural network to be used. The use oftechniques of decision trees, in conjunction withthe management model CRISP-DM, to help in theprevention of bank fraud was evaluated in {5}. Thestudy recognized the fact that it is almostimpossible to eradicate bank fraud and focused onwhat can be done to minimize frauds and preventthem. The research offered a study on decisiontrees, an important concept in the field ofartificial intelligence. The study focused ondiscussing how these trees are able to assist in thedecision making process of identifying frauds bythe analysis of information regarding banktransactions. This information is captured with theuse of techniques and the CRISP-DM managementmodel of data mining in large operationaldatabases logged from internet bank.The Cross Industry Standard Process for Data-Mining – CRISP-DM is a model of a data miningprocess used to solve problems by experts. Themodel identifies the different stages inimplementing a data mining project while, Adecision tree is both a data representing structureand a method used for data mining and machinelearning. [3] describes the Use of neural networksin analyzing the great increase in credit cardtransactions; credit card fraud has becomeincreasingly rampant in recent years. This studyinvestigates the efficacy of applying classificationmodels to credit card fraud detection problems.Three different classification methods, i.e.decision tree, neural networks and logisticregression were tested for their applicability infraud detections. The paper provides a usefulframework to choose the best model to recognizethe credit card fraud risk. Detecting credit cardfraud is a difficult task when using normalprocedures, so the development of the credit cardfraud detection model has become of significance,whether in the academic or business communityrecently. These models are mostly statistics-drivenor artificial intelligent-based, which have thetheoretical advantages in not imposing arbitraryassumptions on the input variables.Other models that have been developed inresearch settings that have promising potential forreal world applications include the CustomerRelationship Model, Bankruptcy Prediction Model,Inventory Management Model, and Financial MarketModel.8

Afr J Comp & ICTVol 4. No. 3. Issue 2Adeyiga et al – A Neural Network Based Model for Detecting Irregularities in e-Banking TransactionsTo increase the body of knowledge on this subject,an in-depth examination of important publiclyavailable predictors of fraudulent financialstatements was offered. They tested the value ofthese suggested variables for detection offraudulent financial statements within a matchedpair’s sample. Self-organizing Artificial NeuralNetwork (ANN) AutoNet was used in conjunctionwith standard statistical tools to investigate theusefulness of these publicly available predictors.The study resulted in a model with a highprobability of detecting fraudulent financialstatements on one sample[1][6]. The studyreinforced the validity and efficiency of AutoNet asa research tool and provides additional empiricalevidence regarding the merits of suggested redflags for fraudulent financial statements. [10]Reviews the various factors that lead to fraud inthe Nigerian Banking industry. The problem offraud in our banking system may have someattachment. Therefore, there must be somefactors that may have led to this fraudulent act.The problems are stated below:Bank malpracticesi. Failure to appoint trusted and honest officialas the banks representative in the clearinghouseii. Failure to change representative on regularbasisiii. Failure to provide locked boxes or bags forcarrying cheques to and from the central banksiv. Inadequate training facilities for clearing staffboth in the offices and central bankv. Negligence in checking clearing cheques fromthe banks to avoid a case of possible shortchange of chequeThis study is important due to the variousguidelines that can be drawn in coming up with asolution based on the problems listed above.3. METHODOLOGYWe perceive neural networks as tools that canrecall and learn patterns of behaviour, detectchanges in patterns, and detect fraud in a paymentcard environment. We research to do thefollowing:• Generate/ create a unique network foreach account in the bank• Generate patterns for the network usingthe client history.• Train the network at user defined epochvalue or till there error value isapproximately 0.• Once the network has learned, feeds thecurrent transaction into the network toknow if it is a fraud or not.Components of The Irregularity Detection Systemi. Neural network based detectorii. Databaseiii. Computer System3.1.1 Neural Network Based DetectorThe neural network based detector is amathematical model or computational model basedon biological neural networks, in other words, it isan emulation of biological neural system. A neuralnetwork is a massively parallel distributedprocessor that has a natural propensity for storingexperiential knowledge and making it available foruse. It resembles the brain in two respects [8].1. Knowledge is acquired by the network through alearning process.2. Interconnection strengths known as synapticWeights are used to store the knowledge.3.1.2 DatabaseThe data are typically organized to model relevantaspects of reality. Structured query language willbe used for the data store. This is the region wherethe account statements will be lodged and detailsof all transactions that take place in each accountare stored in the database.All transactions carried out in a customer’s accountare stored in the database and can be retrievedwhen needed. The database serves as theknowledge base to the neural network, thenetwork trains and generates results based on theinformation in the knowledge base.3.1.3 Computer SystemThe Computer system consists of the platform inwhich the irregularity detection system willoperate, neural network cannot exist on its own. Itis usually implemented in computer systems andrelated devices. The computer also serves as theinterface between the neural network baseddetector and the system user. The computerpresents the input data (user query) to the neuralnetwork.9

Afr J Comp & ICTVol 4. No. 3. Issue 2Adeyiga et al – A Neural Network Based Model for Detecting Irregularities in e-Banking Transactions3.2 Layout of the Irregularity DetectionSystemThe simplified activities that take place in theirregularity detection system1. The input data (customer credit/debittransaction) is presented by the interface(computer) to the database (knowledgebase).2. The network is fed by the knowledge base.3. The network trains based on the data inthe knowledge base4. If the network based detector detects anyirregularity, the transaction is disallowedelse the transaction is allowed. Whateverdecision the network based detectorarrives at is stored at the knowledge base(database)Figure 2. Proposed System Model3.3 The Network Training ModelModeling was done using the unified modelinglanguage (UML). The system is modeled to be ableto look through a client’s transaction history andgenerate an intelligent pattern which will be usedas standard for testing the next transaction.Figure 1. Proposed System componentsFor example consideri. A client that has made withdrawals about1000 times from a particular branch of abank is suddenly withdrawing from afarther (another) branch.ii. A client transaction history can denote hismaximum withdrawal/Deposit to be 5000naira, and suddenly he/shewithdraws/deposits 5,000,000 naira.iii. A client withdrawal/deposit mode andtransaction date difference can also beused as a pointer in identifyingirregularities.In training the multilayer perceptron network(MLP), five different entries will be inputted whichare as follows;Entry 1 : difference between the currentand last transaction based on the number of days.Entry 2: percentage of last withdrawal/deposit tocurrent transaction10

Afr J Comp & ICTVol 4. No. 3. Issue 2Adeyiga et al – A Neural Network Based Model for Detecting Irregularities in e-Banking TransactionsEntry 3: the branch the transaction is taking place(location)Entry 4: patron that is; self or different person?Entry 5: transaction date.The network produces reasonable outputs forinputs it has not been taught how to deal with. Thevarious entries are presented to the network inform of numeric values. Transaction mode forwithdrawal can only be either cheque/slip, ATM,Credit Card. These values can only be presentedwith numeric values in order to generate a pattern.Fixed numbers are chosen at random with thecondition of being centralized around zero. Thenumbers are chosen in order to have distinctiveoutput when presented to the network. Theserandom values assist in reducing the learning timeof the MLP else the network might take a longertime to learn (training a Neural Network involvestrial an error till the result generated matches thetarget output).The same process is followed for branches thougheach branch a bank is uniquely identified by codes.Random values between -1 and +1 are used torepresent the parameters. MLP learns well withvalues cantered on zero. The amount involved inthe transaction is represented as the percent ofthe last transaction amount to the currenttransaction amount; the date is captured by thesystem based on the system’s calendar and clock.The training process and algorithm is thusdescribed below;3.4 The Training AlgorithmSupervised Back propagation training algorithm wasused to train the neural network because of itseffectiveness towards pattern recognition. Trainingset is a collection of training samples gathered. Atraining sample is a pair of input vector plus adesired output value (0.8 or -0.8). The networkwas provided with the training set and allowing itto learn by adjusting weights of its synapses byback propagating the error calculated as thedisparity between the output neuron to theexpected/target value.3.5 The Actual Algorithm1. Identify number of input neurons (same asnumber of elements in the input vector)2. Identify number of hidden layers andnumber of neurons on each layer.(Minimum required for back propagation isone hidden layer but for faster training weused two hidden layers each with 10neurons each).3. Identify number of output neurons. Weused one neuron on the output layerbecause we have one target value.4. Initialize random weight values for thesynapses connecting each neurons ofpreceding layer to the next layer. Withback propagation, weight values should berestricted to between -0.5 and +0.5.5. Choose a random training set from thetraining sample and assign input vector tothe input neurons.6. Propagate all neurons in the forwarddirection to obtain output at the outputlayer.a) The output of each neuron is afunction of its inputs. In particular,the output of the jth neuron in anylayer is described by two sets ofequations on the right: Uj = ∑ (Xi .wij)…………………….(1)b) For every neuron, j, in a layer,each of the i inputs, X i , to that Yj =Fth (Uj + tj) layer is multiplied by apreviously established weight, w ij .These are all summed together,resulting in the internal value ofthis operation, U j . This value isthen biased by a previouslyestablished threshold value, t j , andsent through an activationfunction, F th (tanh function).c) The resulting output, Y j , is an inputto the next layer or it is a responseto the neural network if it is thelast layer.7. Evaluate error values at the output neuronas the difference between obtained outputand the desired output of the training setchosen.8. Backpropagate the error, all the way up tothe input layer.a. Back propagation starts from theoutput layer with the followingequation.Ѡij =ԝij + LR . ej .Xi………………………….(2)11

Afr J Comp & ICTVol 4. No. 3. Issue 2Adeyiga et al – A Neural Network Based Model for Detecting Irregularities in e-Banking Transactionsb. For the input of neuron in theoutput layer, the weight w isadjusted by adding to the previousweight value, w' ij , a termdetermined by the product of alearning rate, LR, an error term,e j , and the value of the input, X i .The error term, e j , for the outputneuron is determined by theproduct of the actual output, Y j ,its complement, 1 - Y j , and thedifference between the desiredoutput, d j , and the actual output.Ej = Yj . (1 - Yj) . (dj - Yj)………………………..(3)9. Calculate and update weight values for allsynapses such that the sum squared valueof the error are minimized.a. Once the error term is computedand weights are adjusted for theoutput layer, the value is recordedand the next layer back isadjusted. A revised weightadjustment process was adoptedfor updating the weights followingthe Equation below.Ⱳij = w’ij + (1 - M) ˑLR ˑ ej ˑ Xj + M ˑ (w'ij -wʹʹij)….(4)Momentum (M) basically allows a change to theweights to persist for a number of adjustmentcycles. The magnitude of the persistence iscontrolled by the momentum factor. If themomentum factor is set to 0, then the equationreduces to that used to adjust the weight of theoutput layer. If the momentum factor is increasedfrom 0, increasingly greater persistence ofprevious adjustments is allowed in modifying thecurrent adjustment. This can improve the learningrate in some situations, by helping to smooth outunusual conditions in the training set.12. Stopping criterion is until the errorobtained at the output layer is at anacceptable value. (0.02)3.6 The Transfer Functions of the NeuralNetwork.Activation function is used to obtain output fromeach neuron. Tangential Activation function wasadopted.The equation is Y=tanh(x).Input nodes shall be presented which are expectedto yield a target output.If the output is not correct, the weights areadjusted according to the formula:w new = w old + α(desired – output)*input……………(6)where α is the learning rate (Cheung and Cannons,2002)Using tanh activation function on all neurons thenetwork, the output of each neuron rangesbetween -1 and +1. Some training set in thetraining sample has a target value of 0.8 and some-0.8.Once the error at the output layer is at anacceptable level, the test data is fed into thenetwork from the input layer and the output valueis derived. The sensitivity bar ranges from -0.8 to0.8 although presented to range from 0 to 100%.The transaction is committed if the output of thetest data is greater than the value of the sensitivitybar, otherwise rolled back.b. The error term is generated by aslightly modified version ofEquation in steps 9 above. Thismodification is:ej = Yj ˑ (1 - Yj) ˑ ∑ (℮k ˑ Ⱳʹjk)…………………….(5)10. Choose another random training set formthe training sample and repeat the stepsabove.11. Train all training set in the training samplein a random selection order. A cyclethrough the training sample is called anepoch.Figure 3. An illustration of the training sets forthe multilayer perceptron network.Source: (Werbos [16]; Rumelhart [14])12

Afr J Comp & ICTVol 4. No. 3. Issue 2Adeyiga et al – A Neural Network Based Model for Detecting Irregularities in e-Banking Transactions3.7 Training and VerificationThe training set is not fixed based on all previoustransactions with the bank. The training set(initials training set) is 70% of the previoustransactions while the testing set (cross validationin the training process) is 30% of the transactions.Cross validation patterns are a fraction of theclient history not used for training rather to checkif the network has learned, the error ratio (crossvalidation error) is shown before training.The function of the current transaction is thereforeto test against the neural network to determine itsperformance.A sensitivity bar will be included so thattransactions with values heading towards acceptedlevel that should be suspicious will not besuccessful. The training set is the set of all knownsamples is broken into two orthogonal(independent) sets:1. Training set 1: A group of samples used totrain the neural network2. Testing set 2: A group of samples used totest the performance of the neuralnetwork. It is also used to estimate theerror rateVerification;i. Provides an unbiased test of the quality ofthe networkii. Common error is to “test” the neuralnetwork using the same samples that wereused to train the neural network.Based on the testing set;1. The network is optimized based on thesesamples, and will obviously perform wellon them.2. Verification doesn’t give any indication asto how well the network will be able toclassify inputs that weren’t in the trainingsetEpochThis is basically the number of times the neuralnetwork should train for a particular transaction.i. One iteration through the process ofproviding the network with an input andupdating the network's weights.ii. The amount of epochs needed for theneural network is not fixed. Typically manyepochs are required to train the neuralnetwork.Irregularities detected by the system based on thetraining set provided to it will be made known tothe bank official, it then becomes the duty of thebank officials to decide if such transactions shouldcontinue or be discontinued.4. CONCLUSIONIn this work, the irregularity detection systemModel has sought to reduce the risk level offraudulent transactions that take place in theNigerian banking industry thereby aiding in thedecrement of bank fraud. This will brings aboutreduced fraudulent transactions if implementedproperly. Neural network technology is appropriatein detecting fraudulent transactions because of itsability to learn and remember the characteristicsof the fraudulent transactions and apply that“knowledge” when assessing new transactions.REFERENCES[1] Idowu, A., “An Assessment of Fraud and itsManagement in Nigeria Commercial Banks”.European Journal of Social Sciences. Vol. 10, No4.,2009, pp 628-640,[2] Adeyemi, k,“Banking Sector consolidation inNigeria: issues and challenges”,2005. Retrievedfrom www.unionbankng.com[3] Aihua Shen, Rencheng Tong, Yaochen Deng,(2007). “Application of classification Models onCredit Card fraud detection.” Service Systemsand Service Management, 2007 InternationalConference on 07/2007; DOI:10.1109/ICSSSM.2007.4280163[4] Amit Khajanchi.. “Artificial Neural Networks:The next intelligence”,2003,. Available atwww.globalriskguard.com/resources/market/NN.pdf[5] Bruno Carneiro, Rafael Sousa, “IdentifyingBanks Fraud Using CRISP-DM and DecisionTrees.” International Journal of ComputerScience and Information Technology (IJCSIT)Vol.2, No.5, October 2010.[6] Fanning. K, Cogger. K. O and Srivastava. R.“Detection of management fraud: A neuralnetwork approach” International Journal ofIntelligent Systems in Accounting, Finance andManagement .1995. 4 113-126.[7] Fawcett, T. and Provost, F. “Adaptive frauddetection”. Data Mining and knowledgeDiscovery,1997. 1:291-316[8] Haykin, S , “Neural Networks; A ComprehensiveFoundation”, 2nd ed. (Englewood Cliffs, NJ:Prentice-Hall). 1999.13

Afr J Comp & ICTVol 4. No. 3. Issue 2Adeyiga et al – A Neural Network Based Model for Detecting Irregularities in e-Banking Transactions[9] Idolor Eseoghene Joseph .”BANK FRAUDS INNIGERIA”; underlying causes, effects andpossible remedies”. African Journal ofAccounting, Economics, Finance and BankingResearch Vol. 6.No. 6. 2010.[10] Ivor Ogidefa . “Fraud in Banking System inNigeria”, November 19, 2008. Available athttp://socyberty.com/law/fraud-in-bankingsystem-in-nigeria/[11] Lemo: T ,”Regulatory Oversight andStakeholder Protection” A paper presented atthe BGL Mergers and Acquisition InteractiveSeminar, Held at Eko hotels $ Suits. V.I, onJune 24.2005.[12] Micheal, H., Diane L, Jose C.P and Don X.“Detecting fraud in the real World”.Availableatstat.belllabs.com/cm/ms/departments/sia/doc/HMDS.pdf. 2001.pp. 7 – 12[13] Cheung and Cannon “An Introduction to NeuralNetworks”, Signal and Data CompressionLaboratory Group Meeting, University ofManitoba, May 27, 2002.[14] Rumelhart, D. E., G. E. Hinton, and R. J.Williams: “Learning internal representations byerror propagation,” in D. E. Rumelhart and J.L. McCleland, eds. (Cambridge, MA: MIT Press),vol. 1, Chapter 8, page 318-62, 1986.[15] Soludo, C., “Consolidating the Nigerian bankingindustry to meet the development challengesof the 21 st century”. Being an addressdelivered to the Special meeting of bankerscommittee held on july 6, 2004 at the CBNheadquarters Abuja.[16] Werbos, P.J.,“Beyond Regression: new tools forprediction and analysis in the behaviouralsciences, “ Ph.D Thesis, Havard University,Cambridge, MA.197414