Neural network algorithms for fraud detection: a comparison of the complementary techniques in the last five years

Journal of Management Information and Decision Sciences Volume 24, Special Issue 1, 2021 1 1532-5806-24-S1-94 Citation Information: Navarrete, A.C. & Gallegos, A.C. (2020). Neural Network Algorithms for Fraud Detection: A Comparison of the Complementary Techniques in the Last Five Years. Journal of Management Information and Decision Sciences, 24(S1), 1-16. NEURAL NETWORK ALGORITHMS FOR FRAUD DETECTION: A COMPARISON OF THE COMPLEMENTARY TECHNIQUES IN THE LAST FIVE YEARS Alberto Clavería Navarrete, Universidad de Sevilla Amalia Carrasco Gallegos, Universidad de Sevilla ABSTRACT Purpose: The purpose of this research is to analyse the complementary updates and techniques in the optimization of the results of neural network algorithms (NNA) in order to detect financial fraud, providing a comparison of the trend, addressed field and efficiency of the models developed in current research. Design/Methodology/Approach: The author performed a qualitative study where a compilation and selection of literature was carried out, in terms of defining the conceptual analysis, database and search strategy, consequently selecting 32 documents. Subsequently, the comparative analysis was carried out, in turn being able to determine the most used and efficient complementary technique in the last five years. Findings: The results of the comparative analysis depicted that in 2019 there was a greater impact of research based on NNA with 11 studies. 27 complementary updates and techniques were identified related to NNA, where deep neural network algorithms (DNN), convolutional neural network (CNN) and SMOTE neural network. Finally, the evaluation of effectiveness in the collected techniques achieved an average accuracy ranging between 79% and 98.74% with an overall accuracy value of 91.32%. Originality/Value: Being a technique which is applied and compared in diverse studies, ANNs uses a wide range of mechanisms concerning training and classification of data. According to the findings of this research, the complementary techniques contribute to the progress and optimization of algorithms regarding financial fraud detection, having a high degree of effectiveness concerning on-line and credit card fraud. Keywords: Neural Networks, Fraud Detection, Algorithm, Finance and Financial Fraud. INTRODUCTION Financial fraud has become relevant in the digital era due to the significant increment of these criminal acts as a result of the available technological advances (Pérez, 2018). The actions taken against fraud are focused on two points: fraud prevention and fraud detection. Fraud prevention tries to block fraudulent transactions, whereas fraud detection relates to successful fraudulent transactions which are identified after being committed (Sandoval, 2019). In fraud detection, the inspection of financial statements requires conventional audit techniques in order to prevent the consequences of fraud, such techniques must distinguish between abnormal data and authentic data, therefore allowing early decision making and strategies that diminish the impact of fraud. Currently, new technologies are being used which are capable of processing large volumes of information, allowing to detect abnormal patterns that would possibly be imperceptible to the human eye (Shwartz, 2014). The development of these technologies utilizes data mining as an exploratory mechanism in the detection, drawing from other methods such as automatic data learning, data bases and artificial intelligence (Mishra et. al, 2013). Journal of Management Information and Decision Sciences Volume 24, Special Issue 1, 2021 2 1532-5806-24-S1-94 Citation Information: Navarrete, A.C. & Gallegos, A.C. (2020). Neural Network Algorithms for Fraud Detection: A Comparison of the Complementary Techniques in the Last Five Years. Journal of Management Information and Decision Sciences, 24(S1), 1-16. In the last decades, complex algorithms have been created for the detection of financial anomalies that allow relating artificial intelligence with fraud detection processes; these algorithms constantly learn to react to situations that are regarded as warning signs, and therefore to early determine an event that may occur in the future (Portela et. al, 2019). Fraud detection is a discipline implemented when prevention mechanisms have failed, maintaining a continuous evolution that overshadows evasive methods adopted by criminals (Pérez, 2018). These fraud detection systems must be cost-effective and efficient, that is to say, the cost of investing in the system must not exceed the loss caused by the fraud, and therefore, it is essential to use models and statistic rules to minimize those costs (Sandoval, 2019). The research will be a documental study that will use bibliographic collection techniques and statistical comparison with the purpose of determining the feasibility concerning neural network algorithms in the detection of financial fraud, since they are considered as one of the most implemented algorithms in the community. A systematic mapping of the bibliography was carried out in order to identify the most implemented and updated components. The result of the bibliographic collection and analysis must answer: Which have been the updates or complements applied to the algorithm? Which has been the implementation and efficacy of the algorithm in the detection of fraud? LITERATURE REVIEW Financial Fraud Detection Methods An intrusion detection system monitors the activity in a server or network, in such a way that it can obtain behavioral patterns, code signings or protocol analyses concerning possible attacks or attempts to breach security (Morales, 2018). The statistical methods for fraud detection vary depending on the size and type of data to be used. The majority of techniques compare the observed data with the expected data (Pérez, 2018). Statistical methods for fraud detection might be supervised and non-supervised. Supervised methodologies utilize methods which are capable of distinguishing fraudulent behaviors from the ones that are not, nevertheless, in their construction, initial reliable data concerning the classification of individuals is used, therefore, this methodology exclusively detects types of fraud which have been previously identified; the system receives data sets with different sample, values and classification parameters, which are processed by a mathematical function that maps an input and output signal, thanks to this, the machine discovers what to do (Cáceres & Velásquez, 2018) (Singh & Jain, 2019). Simultaneously, non-supervised fraud detection methods detect patterns in conjunction with fraudulent components, not only in previous occurrences but also in the detection of new patterns. This is carried out by detecting which individuals deviate from the norm (Rodríguez Pérez, 2018). Initially, these methods have a poor performance, but as they calibrate, their actions improve (Cáceres & Velásquez, 2018) (Singh & Jain, 2019). Finally, mixed learning methods are implemented, which mix supervised and non-supervised mathematical algorithms for the detection of fraud. For its execution, they use a limited amount of previously known and unknown data, mainly used in credit card fraud detection (Singh & Jain, 2019). Artificial Neural Networks Artificial neural networks are similar in their mathematical model to the biological organization and behavior of neurons, which act as a device that is activated when it receives stimuli or disruption in their nerve terminals. These stimuli come from adjacent neurons or nerves. Once the neuron is activated (binary state 1), it begins to stimulate another by means Journal of Management Information and Decision Sciences Volume 24, Special Issue 1, 2021 3 1532-5806-24-S1-94 Citation Information: Navarrete, A.C. & Gallegos, A.C. (2020). Neural Network Algorithms for Fraud Detection: A Comparison of the Complementary Techniques in the Last Five Years. Journal of Management Information and Decision Sciences, 24(S1), 1-16. of extending and progressing towards the nerve terminal of adjacent neurons (Alvarado, 2003) (Rodríguez Pérez, 2018), as visualized in Figure 1. FIGURE 1 REAL NEURAL NETWORK According to biological behavior, neurons have the ability to activate other neurons, causing a chain reaction that, depending on the amount of neurotransmitter fluid, can reduce or increase the initial stimulus (Alvarado, 2003). Neural Network Algorithms (NNA) A neural network probabilistic model is constituted by nodes that act as input, output, or middle processors, linking each node with the next node set by a series of weighted trajectories (Han & Pei, 2014). In financial fraud detection, feed-forward networks with only three layers are used (input, hidden and output). Input stimuli to the neural network are called feature vectors (X1, …, Xn). Weight vector (Wl , …, Wn ) indicates the amount of stimuli that reaches the neuron. Finally, the signal emitted by the output unit (Y1, …, Yn) represents the probability of activity (1) or inactivity (0), used as measurement of criminal behavior suspicion (Alvarado, 2003), as shown in Figure 2. FIGURE 2 ARTIFICIAL NEURAL NETWORK This method is capable of recognizing a complex behavior pattern and classifying each of its regions. Moreover, they are capable of manipulating a number of features that will allow them to perform their functions. Once the neural network is specified, it present a Journal of Management Information and Decision Sciences Volume 24, Special Issue 1, 2021 4 1532-5806-24-S1-94 Citation Information: Navarrete, A.C. & Gallegos, A.C. (2020). Neural Network Algorithms for Fraud Detection: A Comparison of the Complementary Techniques in the Last Five Years. Journal of Management Information and Decision Sciences, 24(S1), 1-16. quicker and more efficient response when analyzing input data (Alvarado, 2003; Rodríguez Pérez, 2018). RESEARCH METHODOLOGY The research stems from a document review, focusing on collecting studies of relevant cases, statistical analysis and data interpretation that supervised fraud detection systems use, by means of neural networks (Arias, 2012; Tamayo & Tamayo, 2003). Using the selected literature, the performance of NNA, its implementation as a method for detecting fraud, the updates or added complements, and analysis of its efficacy were examined. Collection Strategy and Document Selection The literature review concerning financial fraud detection examines the performance of different types of mathematical algorithms with the purpose of obtaining a background knowledge that provides information related with the most relevant ANNs in the last five years. Similarly, complementary techniques in ANN will be studied. The literature review was carried out in three stages. The first stage consisted in a conceptual analysis where keywords were selected for the search string related to the title of this research and according to the researcher’s criteria, using the keywords “Mathematical Algorithms”, “Neural Network”, and “Financial fraud detection”. The keywords were combined and implemented in Spanish and English in order to broaden the search range in the web platform. The second stage is related to the data base used in the research. Google Scholar (http://scholar.google.es/) and Scopus (http://www.scopus.com/scopus/home.url) were used, since they allow Boolean operators in their advanced search tools, implementing keywords and narrowing the search results. Concerning temporal coverage, the literature search ranged from 2015 to 2020 and the type of document included journal articles, papers, reports, and research (Benavent et. al, 2011). Finally, the search strategy was based on a Boolean operator scheme, where the keywords were placed in order to establish the conditions that the documents should meet. With the logical sum operator OR (In Spanish “O”), the keywords, synonyms and relevant concepts to the study were related. By means of the operator AND (In Spanish “Y”), the specific keywords that should appear in a document were established (Benavent et. al, 2011). Figure 3 displays the implemented search scheme. FIGURE 3 SEARCH SCHEME APPLYING BOOLEAN OPERATORS Journal of Management Information and Decision Sciences Volume 24, Special Issue 1, 2021 5 1532-5806-24-S1-94 Citation Information: Navarrete, A.C. & Gallegos, A.C. (2020). Neural Network Algorithms for Fraud Detection: A Comparison of the Complementary Techniques in the Last Five Years. Journal of Management Information and Decision Sciences, 24(S1), 1-16. According to the literature review, from each data base 118 documents were extracted: 87 from Google Scholar and 31 from Scopus. 21 duplicates were excluded as well as 65 documents whose access was restricted. Consequently, documents that meet the following conditions were included:  The title indicated the application of neural networks in cases related to finance.  Documents in Spanish and English  Scientific documents (articles, reports and undergraduate works) Documents with the following conditions were excluded:  They were not primary studies.  They were not related to financial fraud.  They did not carry out a statistical comparison in their algorithm. 32 documents were obtained according to Kitchenham’s, 2003, selection process. The documents were tabulated in Excel, where data concerning the authors, year of publication, main objective of the document, type of addressed problem, neural network models implemented, mathematical complements for the processing of data, and conclusions were extracted. Comparative Analysis The study of ANN was carried out by means of a comparative method, that allowed to describe each of the complementary techniques and therefore, to sustain the feasibility for its application (Pérez, 2007). Similarly, each of the algorithms is compared and analyzed by means of similarity and difference matrixes so as to determine which is the ideal one for the detection of financial fraud. This assessment was performed by means of pair matrixes or prioritization matrixes; these are a tool that allows making decisions using established criteria or conditions to rate the problem so as to determine the best option based on the decisionmaking methodology by Linán (2007). The feasibility and efficacy criteria for the rating were determined based on the literature review and the results provided by it, where the following questions were posed: Does the document clearly explain the algorithm along with the implemented components? Is the study experimental? Does the study give answer to the accuracy parameters of the algorithm? The level of importance of each question will be established based on a rating scale, as shown in Table 1. Table 1 LEVEL OF IMPORTANCE OF THE QUESTIONS Level of Importance Rating Scale Very good information 10 Good information 5 Acceptable information 1 Little information 0,2 Does not fulfill with the information 0,1 The comparison concerning the feasibility criteria by means of the pair matrix will allow calculating the Weighting Factor (WF) and Option Weight (OW) that will be used to determine the final rating of each of the algorithms, according to the procedure posed by Liñán (2007). Journal of Management Information and Decision Sciences Volume 24, Special Issue 1, 2021 6 1532-5806-24-S1-94 Citation Information: Navarrete, A.C. & Gallegos, A.C. (2020). Neural Network Algorithms for Fraud Detection: A Comparison of the Complementary Techniques in the Last Five Years. Journal of Management Information and Decision Sciences, 24(S1), 1-16. Efficacy of the Algorithm When determining the efficacy in the model, the direct or indirect use of the bidimensional matrix was verified in each document, with a binary scheme that displays the behavior of the variables in each analyzed document. For the accuracy assessment when identifying the thematic classes visualized in Figure 4, where the columns are results given by the forecasting of the algorithm, the rows represent the real classification of the subject, while the diagonal of such matrix indicates the subjects that were correctly classified by the model (Sandoval, 2019; Liñán, 2007). FIGURE 4 CONFUSION MATRIX (CHAVÉZ SANDOVAL, 2019) In the previous figure, the four terms to predict indicate the TP; amount of positive predictions that are correctly classified by the model, TN; fraudulent predictions that are correctly classified by the model, FP; number of positive predictions incorrectly classified by the model, and the FN that identify the fraudulent predictions which were incorrectly classified by the model (Sandoval, 2019). From the evaluation of the algorithm by means of the confusion matrix, complementary performance indicators emerge that allow assessing the efficacy and quality of the studied algorithm. The accuracy establishes the quality of the predictions made, evaluating the ability of the model to correctly classify the positive and negative cases (Sandoval, 2019). RESULTS AND DISCUSSION The literature compilation and selection presented a total of 32 documents that implement neural network algorithms for fraud detection. Each document was identified with the reference code D[x], where [x] is the document identifier in the bibliography. In Table 2 the distribution of each document by year can be appreciated. According to the aforementioned table, in 2019 a high tendency towards using the studied algorithm was noticed with 11 documents, whereas in 2015 and 2016 its use was rarely implemented by research, with one and two documents respectively. Table 2 DOCUMENT DISTRIBUTION BY YEAR Year Amount of Documents Document Identification 2020 4 D[1], D[18], D[38] y D[43], 2019 11 D [47], D[24], D[11], D[13], D[4], D[33], D[23], D[48], D[19], D[3] y D[40] Journal of Management Information and Decision Sciences Volume 24, Special Issue 1, 2021 7 1532-5806-24-S1-94 Citation Information: Navarrete, A.C. & Gallegos, A.C. (2020). Neural Network Algorithms for Fraud Detection: A Comparison of the Complementary Techniques in the Last Five Years. Journal of Management Information and Decision Sciences, 24(S1), 1-16. 2018 8 D[6], D[5], D[2], D[32], D[26], D[51], D[46] y D[7] 2017 6 D[30], D[50], D[27], D[12], D[10] y D[14] 2016 1 D[34] 2015 2 D[45] y D[16] Identification of the Boom of Neural Network Techniques Applied to Financial Fraud Detection The documents present studies of the principal Neural Network Algorithm (NNA) and 27 complementary techniques were identified for the optimization of the ANN. In order to answer the posed question concerning which complements or updates have been applied to the algorithm, we begin from the information compiled in Table 3, were the techniques used in each document are displayed. Tabla 3 IDENTIFICATION OF ANN COMPLEMENTARY TECHNIQUES Assigned Variable Complementary Technique Document Identification Y1 ANN with HAA (Harmony Annealing Algorithm) and HSA (Harmony Search Algorithm) D[38] Y2 ANN with D[18] Y3 Recurrent long-short term memory ANN (LSTM ANN) D[43] y D[48] Y4 Bidirectionnal recurrent ANN (BRNN) D[1] Y5 Twin ANN based on convolutional neural networks (CNN) and long-short term memory (LSTM) D[47] Y6 ANN with Denoiser autoencoder (DAE) D[24] y D[4] Y7 Probabilistic Neural Network (PNN) D[11] y D[32], Y8 Multilayer Feedforward Neural Networks D[11] y D[26], Y9 ANN with Genetic algorithm (GA) D[33] y D[16] Y10 ANN with SMOTE D[10], D[33] y D[23] Y11 ANN with data mining D[40] Y12 Deep Neural Network (DNN) D[3], D[6] y D[5], Y13 Deep Neural Network (DNN) with Latent Dirichlet Allocation (LDA) D[6] Y14 Neural network with Monarch Butterfly Optimization (MBO) algorithm D[16] Y15 Convolutional Deep Neural Networks (CDNN) D[2] Y16 Multilayer perceptron ANN (MLP) D[50] y D[30] Y17 ANN with Gradient Descent Adaptive (GDA) learning D[27] Y18 ANN with bayesian regularization learning (BR) D[27] Y19 ANN with bayesian regularization learning (LM) D[27] Y20 ANN with hidden layers D[12] y D[14] Y21 Convolutional ANN based on functional sequencing D[51] Y22 ANN with Radical Basis Functions (RBF) D[7] Y23 ANN with Back Propagation Algorithm (BP) y Particle Swarm Optimization (PSO) D[45] Y24 ANN with Particle Swarm Optimization (PSO) D[16] Y25 ANN with Gradient Boosting Decision Tree (XGBoost) D[13] y D[5] Y26 ANN with logic regression D[19] y D[46] Y27 ANN with Cortical Learning Algorithm (CLA) y hierarchical temporal memory algorithm (HTM) D[34] Journal of Management Information and Decision Sciences Volume 24, Special Issue 1, 2021 8 1532-5806-24-S1-94 Citation Information: Navarrete, A.C. & Gallegos, A.C. (2020). Neural Network Algorithms for Fraud Detection: A Comparison of the Complementary Techniques in the Last Five Years. Journal of Management Information and Decision Sciences, 24(S1), 1-16. The study of ANN has been complemented by several techniques for optimizing its results. The table above showed that the main most studied complementary techniques have been deep neural network algorithms (DNN), convolutional neural networks (CNN) and neural networks with SMOTE. The DNNs present different versions of its implementations in documents D[2], D[6], D[3], y D[5]. In most current documents, specifically in the 2019 compilation; D[3] uses a sequential model based on DNN which is able to spot the sequential pattern of online transactions, drawing from memory networks in order to optimize the efficacy and interpretation of the model data. In 2018, document D[6] proposed a model for fraud detection in automobile insurance where an analysis of the features and description of the text was performed, in turn training the DNN based on such features and thus being able to detect anomalies on the claims. Although the studies found in D[2] and D[3] concerning DNN are studies in different fields of financial fraud, the results obtained in D[5] were more optimal when implementing these techniques on neural networks, since configurations in the parameters and structure of the network of up to 1000 series were made, with a learning ratio that ranges between 0,01 and 0,05; these experiments generated the efficient architecture of the DNN. In the convolutional neural network studies (CNN), D[47], D[2] and D[51] can be found. The most updated research concerning this technique was presented in D[47], where a twin convolutional network to solve data sample imbalance problems of online transactions and implementation of the LSTM structure to generate user data in the memory of the model is displayed. In 2018 online transactions were approached with this technique from another angle in D[51], where the CNN elaborates a sequence layer of input features that induces the reorganization of the patterns of unprocessed transactions in order to generate different convolutional shapes. In the same year D[2] was published, which combined the two aforementioned techniques, where deep learning complements, detailed records of frauds by clients and extraction of learning features and classification of fraudulent events in order to generate a convolutional deep neural network model (CDNN) were used, outperforming other traditional automatic learning algorithms. The third most implemented technique in neural networks is SMOTE, studied in D[10], D[33] and D[23]. In the three documents, SMOTE was used for preliminary data sampling, improving the forecasting, and identifying criminal anomalies. In 2017, D[10] puts SMOTE to test by applying it to different automatic learning methodologies, adding a cost matrix to the treatment of preliminary data. As a result, favorable results were obtained in the variables of the confusion matrix, as well as over 90% accuracy with the neural network algorithm. It is evident that in the last 5 years the updates and complementary techniques in neural network algorithms has been varied, with research sustained by two pieces of research or in particular cases, by innovative models. From this array of applied techniques, the final objective of each study was emphasized, in order to indicate the projection and the best field for further development of the ANN, answering which has been the implementation of the fraud detection algorithm. By synthesizing the data, the aim of the algorithm update was extracted with the complementary technique. In table 4, the specific cases in the fraud detection field towards which the documents were focused on are displayed. Table 4 APPLICATION ON FINANCIAL FRAUD DETECTION Complementary Technique Objective Document Identification Credit Card Fraud Detection D[43], D[1], D[27], D[50], D[12], D[24], D[4], D[33], D[48], D[40], D[34], D[26], D[46], D[7], D[45] y D[16] Online Transactions Fraud Detection D[18], D[10], D[47], D[11], D[13], D[23], D[19], D[3], D[51], D[2], D[32] y D[30] Journal of Management Information and Decision Sciences Volume 24, Special Issue 1, 2021 9 1532-5806-24-S1-94 Citation Information: Navarrete, A.C. & Gallegos, A.C. (2020). Neural Network Algorithms for Fraud Detection: A Comparison of the Complementary Techniques in the Last Five Years. Journal of Management Information and Decision Sciences, 24(S1), 1-16. Inadequate User Behavior Fraud Detection D[38] y D[14] Automobile Insurance Fraud Detection D[6] Online Security Breach Detection D[5] From the 32 compiled documents, 16 address credit card fraud detection, 12 documents address online transaction fraud and the remaining 4 were focused on automobile insurance fraud and inadequate user behavior. Comparative Analysis of the Complementary Techniques for Neural Network Algorithms Concerning the comparative results of efficacy and feasibility in each of the compiled techniques, the following criteria were determined: sustained theoretical background of the algorithm, description and mathematical development of the algorithm, experimental study of the algorithm and accuracy of the algorithm. The sustained theoretical background provides scientific support from previous research that guarantees the reliability and logic of the applied technique. The description and mathematical development establish the essential characteristics of the algorithm. The experimental study proves the verification of the algorithm implemented in the studies and the accuracy assesses the result of the algorithm in financial fraud detection. In table V, the data matrix with the studied complementary techniques and the criteria to analyze are displayed. Such criteria are represented in the following fashion: X1 (sustained theoretical background of the algorithm), X2 (description and mathematical development of the algorithm), X3 (experimental study of the algorithm) and X4 (accuracy of the algorithm). The results shown in table 5 indicate the assessment of the criteria in each applied technique according to the previously mentioned rating scale. According to the criteria and rating scale, the displayed documents will present a total sum of over 20 points, considering it as research that provides results classified as “Good Information”. Table 5 EFFICACY AND FEASIBILITY CRITERIA OF THE NEURAL NETWORK COMPLEMENTARY TECHNIQUES (DIFFERENCE AND SIMILARITY MATRIX) Assigned variable X1 X2 X3 X4 ∑ Total Y1 5 1 5 10 21 Y2 5 10 10 10 35 Y3 10 5 5 5 25 Y4 5 1 5 10 21 Y5 5 1 5 5 16 Y6 10 5 10 10 35 Y7 1 5 5 5 16 Y8 5 1 5 5 16 Y9 1 5 5 5 16 Y10 10 5 10 10 35 Y11 1 0.1 0.2 5 6.3 Y12 5 5 10 10 30 Y13 5 1 5 5 16 Y14 1 5 5 10 21 Y15 5 1 5 5 16 Journal of Management Information and Decision Sciences Volume 24, Special Issue 1, 2021 16 1532-5806-24-S1-94 Citation Information: Navarrete, A.C. & Gallegos, A.C. (2020). Neural Network Algorithms for Fraud Detection: A Comparison of the Complementary Techniques in the Last Five Years. Journal of Management Information and Decision Sciences, 24(S1), 1-16. Vinayakumar, R., Ganesh, H.B., Prabaharan, P., Anand, K.M. & Soman, K.P. (2018). Deep-net: Deep neural network for cyber security use cases. Xinxin, Z., Zhaohui, Z., Lizhi, W. & Pengwei, W. (2019). A model based on the Siamese neural network for the detection of fraud in online transactions. International Joint Conference on Neural Networks (IJCNN). Yibo, W. & Wei, X. (2018). Leveraging deep learning with LDA-based text analytics to detect automobile insurance fraud. Retrieved from https://doi.org/10.1016/j.dss.2017.11.001. Yogesh, P., Karim, O., Vassil, V., & Jun, L. (2019). Remote bank fraud detection framework using sequence learners. Internet banking and commerce magazine. Zanin, M., Romance, M., Moral, S. & Criado, R. (2017). Credit card fraud detection through parenclitic network analysis. Complexity, 48. Zhaohui, Z., Xinxin, Z., Xiaobo, Z., Lizhi, W. & Pengwei, W. (2018). A model based on convolutional neural network for online transaction fraud detection. Retrieved from https://doi.org/10.1155/2018/5680264.