=Paper=
{{Paper
|id=Vol-1981/paper1
|storemode=property
|title=Survey of Common Design Approaches in AML Software Development
|pdfUrl=https://ceur-ws.org/Vol-1981/paper1.pdf
|volume=Vol-1981
|authors=Alexander Semenov,Artem Mazeev,Dmitry Doropheev,Timur Yusubaliev
}}
==Survey of Common Design Approaches in AML Software Development==
https://ceur-ws.org/Vol-1981/paper1.pdf
Survey of Common Design Approaches in AML Software
Development
Alexander Semenov Artem Mazeev
JSC NICEVT JSC NICEVT
Moscow, Russia Moscow Institute of Physics and Technology
alxdr.semenov@gmail.com Moscow, Russia
mav367@mail.ru
Dmitry Doropheev Timur Yusubaliev
Moscow Institute of Physics and Technology Quality Software Solutions ltd
Moscow, Russia Moscow, Russia
dmitry@dorofeev.su ytr@kpr-it.com
Abstract The role of financial institutions is to find ways to
identify, among the huge number of operations that oc-
In the recent years, money laundering activ- cur every day, those suspicious transactions and then
ity has become more and more ubiquitious investigate them in more detail [3]. There are many
all over the world. In the paper we survey illegal money laundering schemes of complex nature,
the technical aspects of anti-money laundering and new sophisticated schemes appear exploiting ever
systems (AML). We briefly present the prin- increasing richness of money forms and economic ac-
ciples of money laundering process and fea- tivity [4]. Some schemes are recognized by FATF (Fi-
tures of the illegal activity that can arise in the nancial Action Task Force inter-governmental organi-
graph of money transactions, company or user zation that aims to set standards and promote effective
profiles. Then we present a detailed analysis implementation of legal, regulatory and operational
of anomaly detection, machine learning and measures for combating money laundering, terrorist
neural networks techniques in the context of financing and other related threats to the integrity of
AML systems. the international financial system [5]).
Keywords: anti-money laundering, machine There are a large number of papers that consider
learning, anomaly detection, graph mining the solution of this problem, but most of them be-
long to authors from China [6], Australia, India [7],
Sweden, [8], Poland, Ireland [9], Egypt [10], Pakistan
1 Introduction
[11], Saudia. IT Giants and authors from the most
Money laundering (ML) is a criminal activity or pro- advanced natures do not publish papers which con-
cess that deals with criminal proceeds to disguise their tain technical considerations of anti-money laundering
illegal origin and make them appear legitimate. In (AML) systems with corresponding keywords.
recent years, money laundering activity is becoming There exist only small number of survey articles
more and more ubiquitious all over the world [1]. Be- on AML systems and algorithms [12, 13, 14]. Related
tween 1.5 trillion USD and 2.8 trillion USD or between works are cited in [11,15,16,17] and some other papers.
2% and 5% of global gross domestic product (GDP) is Besides money laundering there is another type of
lost annually through money laundering worldwide [2]. illegal activity. Fraud is a crime where the objective is
to gain money by an illegal form [18]. Fraud detection
Copyright c by the paper’s authors. Copying permitted for
is similar to money laundering but our survey considers
private and academic purposes.
only money laundering.
In: V. Voevodin, A. Simonov (eds.): Proceedings of the
GraphHPC-2017 Conference, Moscow State University, Russia, The survey is an introduction to technical and cor-
02-03-2017, published at http://ceur-ws.org. responding aspects of anti-money laundering systems.
1
The organization of the survey is as follows. Sec- However, there are also two-stage and four-stage
tion 2 deals with principles of money laundering pro- models [20]. In the four-stage model placement oc-
cess and features of the illegal activity that can be curs simultaneously in many places (small amounts for
found in the graph of money transactions, company many different accounts, often through acquaintances,
or user profiles. Section 3 provides a historical re- relatives).
view of main anti-money laundering approaches. Sec- The financing of terrorism differs from money laun-
tion 4 presents detailed analysis of the state-of-the-art dering in that the funding sources can be of legal ori-
techniques and AML systems. Finally, we conclude in gin.
Section 5 . It is clear from the three-stage model that a typical
money laundering scheme involves multiple transac-
tions, conducted through a variety of different chan-
2 Money Laundering Features
nels, banks, by a group of parties (individuals, busi-
The process of money laundering usually consists of nesses, etc.).
three stages [19]: Different schemes can be borrowed from the rules
of committees that exist in different states. For ex-
• Placement is a transfer of cash into financial sys- ample, in USA [21], in Australia [22]. There are also
tem. Placement is carried out in credit institu- state-controlled organizations that help businesses, for
tions, security markets, retail trading. The goal example in the USA — FINRA [23]. Some typical
of placement is to conceal or disguise source or money laundering schemes are described in [19, 24].
ownership of the illegal funds. Often funds are Basic characteristics of a money laundering scheme
placed in the foreign countries. The stage of place- are the following: a role assignment between the par-
ment of criminal proceeds is the most unreliable ties, a particular execution order of the transactions,
stage in the process of money laundering. At this synchronization of transactions as per time and the
stage there is the best chance to identify illegally amount of transactions, etc.
received funds. The authors of [25] note the following important
• Layering is a separation of criminal proceeds features for AML:
from their sources through complex chains of fi-
nancial transactions aimed at disguising the trace • Type of the main activity of the company (retail,
of the illegal funds. Various financial operations gambling, intermediary services);
• Type, amount and size of transactions (cash re-
overlap one another in order to complicate the
ceipt, withdrawal of cash from ATM);
work of AML systems and experts aimed to iden-
• Geographical factor (offshore operation, partici-
tifying criminal proceeds and persons who legal-
pation in the transaction of the company or a
ize them. If the placement of criminal funds was
bank with a critically low level of AML);
successful, that is, it was not discovered, then it
• Specificity of the organizational structure (for ex-
becomes much more difficult to detect illegal op-
ample, the company with too young founders);
erations at this stage. • Events preceding the transaction (changes in the
• Integration is the last stage of the legalization ownership of the company, obtaining a large loan);
process, directly aimed at creating the appear- • Transaction history.
ance of legality to the criminally obtained capital.
Funds separated in the previous stages are con- 2.1 Client Profiling
solidated at the integration stage into some form,
which is convenient for the customer: money on In AML it is important to determine potential risky
the account in a first-class bank, liquid securities, parties (users, clients). AML systems can have an in-
real property assets. The laundered funds are in- creased focus on risky users.
vested further in the legal sectors of the econ- The following can be used as features of risky users
omy, which also creates a basis for new crimes. [3, 26]:
To simplify further legalization of funds, crimi- • Large deposits and withdrawals;
nal communities acquire credit organizations and • Periodicity of transactions;
other financial institutions, and also buy signifi- • Transactions in risky areas regarding money laun-
cant shares in the ownership of enterprises in the dering (in free trade-industrial zones);
real sector of the economy, transferring them un- • Transactions of less than a specified threshold
der their own control. If the money laundering amount in order to avoid control of AML systems;
trace has not been identified in the two previous • Degree of bank service usage;
stages, then it is extremely difficult to separate le- • Reactivation of off-line accounts;
gal funds from illegal ones at the integration stage. • Account age.
2
2.2 Graph Patterns Real graph patterns can be highly complicated. In
the paper [28] an algorithm for searching of “volca-
Money laundering can be detected in the graph of
noes” and “black holes” patterns is proposed, which
money transactions. Relations between parties in the
actually represents the placement, layering and inte-
graph of money transactions produce social network.
gration model. Figure 4 shows these patterns.
The graph mining methods and methods of social net-
works analysis can be applied to detect suspicious
transactions [27].
There are different patterns (templates) in the
transaction graph that can be suspicious. A com-
mon structure of a subgraph is in Figure 1 and can
correspond to the placement, layering and integration
model. Initially money is placed in X vertex. Two
other typical AML subgraphs are shown in Figures
2 and 3. It refers to cyclic business activity to con-
ceal money’s genesis. Of course, these examples are
only schematic and real subgraphs can be different and Figure 4: ”Volcanoes” and ”black holes” in the trans-
complicated. Many relevant papers solve the problem action graph.
of searching of clusters in the graph by some criteria.
3 Anti-Money Laundering Principles
The AML problem can be solved by using different
approaches:
• Rule-based approach;
• Anomaly detection;
• Machine learning.
One of the first AML-systems [29] was created in
1995 and was based on rules. Rules can be very com-
plex and can be defined with use of decision trees [30].
Rules are formulated by experts and can very accu-
Figure 1: Common structure of a subgraph in the rately detect criminal schemes. However, this tech-
placement, layering and integration model. nique is human dependent, not flexible and not auto-
matic. Further, it can not be used to recognize new
typologies of fraudulent transactions.
Data mining techniques involve interdisciplinary
methods from machine learning, statistics and
databases. Anomaly detection [31, 32, 33] is the
one of the data mining problems. Systems that use
the anomaly detection began to appear after systems
based on rules. The report [34] provides an overview
of data mining techniques for detecting fraud in differ-
ent areas. The advantages of anomaly detection are a
capability of discovering new laundering patterns and
Figure 2: Common structure of a subgraph represent-
an ability to customize normal activities.
ing illegal cyclic business activity.
Machine learning and artificial intelligence are
big topics in the financial services sector these days.
Machine learning is increasingly used in the modern
systems [15, 35]. The most popular problem is the
classification problem (two-class classification). The
goal is to define whether the given transaction is sus-
picious or not after procedure of training on prece-
Figure 3: AML subgraph pattern. dents. The test set of precedents specifies information,
which transactions are suspicious and which are not.
3
The problem can be considered in terms of probabil- is to find unusual points (outliers). Outliers can be
ity, when for each transaction a probability (weight) of searched using different algorithms [40], for example,
transaction being suspicious is assigned. The Bayesian using the K-nearest neighbors algorithm [41].
classification methods can be used for the problem. It is also possible to use statistical methods to
This type of detection is only able to detect frauds anomaly detection [31]. One needs to build distribu-
similar to those which have occurred previously and tion function of some feature as a random variable,
have been classified by a human. To detect a novel and then select the parameters, by which the outliers
type of fraud it may require the use of an unsupervised can be detected.
machine learning algorithm. Despite the absence of the training phase, it is pos-
Earlier AML-systems often used the principle of sible to develop some method of determining the sus-
analyzing the actions of a particular person, but in picious transactions to suggest the user, which trans-
modern AML-systems, the analysis of the network (or actions may be suspicious. For example, a good simple
graph) of transactions is increasingly used. For exam- idea is to use the knowledge about customers and de-
ple, in [36] on the basis of bank data, a social network tect activities that are not typical for reliable clients
is built between the clients of the bank, which are con- [42].
nected by transactions. The clustering based techniques group customers
The papers [25, 37] consider a search of patterns that perform similar kind of transactions into a single
in graphs (template subgraphs) that are most similar cluster and then categorize either small-size clusters or
to the known patterns. The principle of approximate outliers as anomalous [11].
search is applied when there is no need for a fixed Theoretically, supervised anomaly detection
pattern matching, which means that the search is more methods can provide better results than unsupervised
flexible as it allows some deviation from the templates. methods, since they are based on more information.
However, the training sets usually contain some noises
4 Modern Techniques and Systems that result in higher false alarm rates, and obtaining
It is hard to strictly classify approaches to the devel- accurate training set is a difficult problem [33].
opment of AML systems. Each AML subsystem can
uses different data mining techniques. 4.2 Graph Mining
Many techniques have been developed in the past
4.1 Anomaly Detection decades for spotting outliers and anomalies in unstruc-
Anomaly detection is an identification of elements of tured collections of multi-dimensional data points.
a given dataset that do not conform to the expected On the other hand, data objects cannot always be
pattern. These anomalies occur very infrequently but treated as points lying in a multi-dimensional space
may signify a large and significant threat such as cyber independently [43]. Transaction graph is a good
intrusions or fraud, i.e. it is necessary to detect the abstraction for data represeting in AML systems,
anomaly (suspicious) data in the background of the inter-dependencies between transactions should be ac-
other data. counted for during the anomaly detection process.
The advantages of anomaly detection are capability In [43] a comprehensive survey of graph based
of discovering new laundering patterns and customiza- anomaly detection techniques is presented.
tion ability of normal activities. The anomaly detec- Graph pattern matching or graph isomorphism can
tion drawback is a high rate of false alarms. be appropriate technique for searching abnormal sub-
The unsupervised anomaly detection techniques graphs in the transaction graph. But exact matching
do not need training data. This kind of methods is makes too strict requirements, and inexact algorithms
beneficial when some abnormal behavior has not been should be used instead. For example, in the paper [28]
demonstrated in the training sample, for example, if an algorithm for searching of “volcanoes” and “black
the sample is small. holes” patterns is proposed, which actually represent
There are a large number of methods and algo- the placement, layering and integration model. Figure
rithms for anomaly detection [32], for example: clus- 4 shows these patterns.
tering techniques [38], unsupervised neural network, Database information about attendees and their re-
fuzzy C-means [39]. lations can be useful for anti-money laundering algo-
A set of the features is necessary for algorithms of rithms. So social network analysis can be additional
anomaly detection as well as for supervised learning. technique that can be used in AML systems [44]. For
Note that the vector of features (numbers) can be rep- example, in [36] on the basis of bank data, a social
resented as a point in n-dimensional space (where n network is built between the clients of the bank, which
is the number of features). The goal of the algorithm are connected by transactions. Features are calculated
4
for each customer (betweenness centrality, page rank, 4.5.1 Australian AML system
etc.). Based on the features, each client is assigned a
The Australian AML system developed for the Aus-
role, for example, it can be an isolator role that isolates
tralian Transaction Reports and Analysis Centre
a certain group of clients from the rest; Another role
(AUSTRAC) [49] is described in the recent papers
– the role of a communicator, which, on the contrary,
[15, 50] in 2016, 2017.
connects several groups of customers. Further analy-
sis of new transactions checks whether the previously The system analyzes two types of transactions:
identified role of the client is consistent with its new large cash deposits and oversea transfers. The sub-
behavior. ject of the analysis is a graph in which parties are con-
nected by edges of two types: one type for transaction,
4.3 Fuzzy Graph Patterns another type for suspicious relations between parties.
There can be a lot of suspicious relation cases, for ex-
In modern AML systems, the search of subgraphs that ample, two persons own the same account. Transac-
approximately satisfy the given criteria is an impor- tion edge has a weight.
tant component. The problem is called fuzzy graph The system performs the following actions:
isomorphism. In [25] authors present an algorithm for
finding a fuzzy set of maximal cliques, they claim that 1. Modeling of relationships between parties using a
method based on the algorithm can contribute to mod- graph with attributes;
ern AML systems that use comprehensive information
from various information sources about actors as well 2. Community extraction from the transaction
as experts evaluation. graph;
In [37] authors propose a method for mining tran-
scation graphs based on building a model that is 3. Calculation of the features from the extracted
parametrized by fuzzy numbers. These numbers repre- communities;
sent parameters of transactions and of the transaction
subgraphs to be detected. The method uses genetic 4. Supervised learning.
algorithm for parameters optimization.
Fig. 5 shows the general scheme of the system.
4.4 Client Profiling Systems
It has been noted that in AML systems it is important
to determine potential risky parties (users, clients).
In [3] a client profiling subsystem is proposed. All
bank clients are clusterized with use of k-means [45]
and other algorithms by some features. Unsupervised
anomaly detection techniques are used for discover-
ing new patterns with subsequent rule generation al-
gorithm [46].
In [26] fuzzy rules are used for description genera-
tion of risky users. Fuzzy rules retain the advantages
of fuzzy expert systems, while reducing the need for an
expert. The 5-layered neural network trains and im-
proves the fuzzy rules. The training method includes
a combination of least squares and back propagation.
4.5 Machine Learning
In recent years, machine learning and artificial intel-
Figure 5: Scheme of the Australian AML system.
ligence have seen an increasing interest and popular-
ity in the financial services community [15, 37, 47, 48].
Machine learning is a particularly powerful tool for The development of algorithms for community de-
prediction purposes. Supervised methods (also known tection in a graph with attributes is an important sci-
as classification methods) required a labeled training entific challenge. There are only limited number of
set containing both normal and anomalous samples to concerned papers, for example, the preprint of the pa-
construct the predictive model. per [51]. The existing community detection algorithms
AML system based on machine learning principles can be poorly adapted to a specific task, for example,
is described in the next subsection 4.5.1 . they detect communities of too large size. At the same
5
time, in the problem of detection of suspicious transac- method has the highest detection rate and the lowest
tions the size of the criminal communities are relatively false positive rate.
small. In [26] a 5-layered neural network trains and im-
The k-step neighbors algorithm is used to commu- proves the fuzzy rules, which are used for generation of
nity extraction in the considered system. Authors use risky users description. The training method includes
its own methodology for processing the communities a combination of least squares and back propagation.
that can overlap.
Algorithms of the classification problem in machine 5 Conclusion
learning need a set of formalized features of each ob-
ject. In the considered system different features are In the paper we survey the technical aspects of anti-
calculated for the extracted communities. Features are money laundering systems. We briefly present the
determined by experts (unpublished, closed part of the principles of money laundering process and features of
paper). Authors note that features can be divided into illegal activity that can arise in the graph of money
several categories: transactions, company or user profiles. Then we
present a detailed analysis of anomaly detection, ma-
1. Demography. Aggregated features that describe chine learning and neural networks techniques, and the
persons in the graph, for example, the average modern Australian AML system.
age; Research is being conducted with the finance sup-
port of the Ministry of Education and Science of the
2. Graph. Features that describe the structure of Russian Federation Unique ID for Applied Scientific
the graph; Research (project) RFMEFI57816X0218. The data
presented, the statements made, and the views ex-
3. Transaction. Aggregated features that describe
pressed are solely the responsibility of the authors.
transactions in the extracted communities, for ex-
ample, total sum of transactions;
References
4. Dynamics. Varied over time features. For exam-
[1] Y. Yang, B. Lian, L. Li, and P. Li, “A RBF neu-
ple, the number of rapidly changing transaction
ral network model for anti-money laundering,”
sizes.
Proceedings of the Wavelet Analysis and Pattern
In the system a support vector machine (SVM) Recognition International Conference, pp. 209–
and a random forest is used as machine learning al- 215, 2008.
gorithms. Evaluation of the system gains 74–98 %
[2] B. Unger, “Offshore Activities and Money
precision, the authors claim that the constructed clas-
Laundering: Recent Findings and Challenges,”
sifiers are able to achieve a level of performance that
European Parliament Directorate-General
is suitable for use in a real environment.
for Internal Policies, 2017. http://www.
europarl.europa.eu/RegData/etudes/STUD/
4.6 Neural Networks
2017/595371/IPOL_STU(2017)595371_EN.pdf
It seems that while machine learning techniques are (accessed: 29.09.2017).
sufficient for AML purposes, neural networks seems to
be too powerful tool for AML. But there are several [3] C. Alexandre and J. Balsa, “Client Profiling
examples of using neural networks in AML. for an Anti-Money Laundering System,” CoRR,
In [1] authors propose a radial basis function (RBF) vol. abs/1510.00878, 2015. https://arxiv.org/
neural network model based on clustering algorithm. pdf/1510.00878.pdf (accessed: 29.09.2017).
An RBF network is a three-layer feed-forward neu-
[4] C. Jedrzejek, M. Falkowski, and J. Bak, “Graph
ral network which consists of an input layer, a hid-
Mining for Detection of a Large Class of Finan-
den layer, and an output layer. RBF network uses
cial Crimes,” http://ceur-ws.org/Vol-483/
center vectors as the parameters of the hidden layer.
paper4.pdf (accessed: 29.09.2017).
In the proposed algorithm center vectors are obtained
by clustering algorithm. During the training stage [5] Financial Action Task Force (FATF).
the algorithm clusters the sample data, the centers http://www.fatf-gafi.org/home/ (accessed:
of clusters are center vectors of radial basis function. 29.09.2017).
The authors note that SVM can get good result, but
the time cost is high. The proposed method is com- [6] X. Luo, “Suspicious Transaction Detection for
pared against support vector machine (SVM) and out- Anti-Money Laundering,” International Jour-
lier detection methods, which show that the proposed nal of Security and Its Applications, vol. 8,
6
no. 2, pp. 157–166, 2014. http://www.sersc. [14] N. A. L. Khac, S. Markos, M. O’Neill,
org/journals/IJSIA/vol8_no2_2014/16.pdf A. Brabazon, and M.-T. Kechadi, “An inves-
(accessed: 29.09.2017). tigation into Data Mining approaches for Anti
Money Laundering,” International Conference on
[7] C. Suresh, K. T. Reddy, and N. Sweta, “A Hy- Computer Engineering and Applications, vol. 2,
brid Approach for Detecting Suspicious Accounts 2009. http://www.ipcsit.com/vol2/94-C140.
in Money Laundering Using Data Mining Tech- pdf (accessed: 29.09.2017).
niques,” Information Technology and Computer
Science, pp. 37–43, 2016. [15] D. Savage, Q. Wang, X. Zhang, P. Chou,
and X. Yu, “Detection of Money Laundering
[8] E. A. Lopez-Rojas and S. Axelsson, “Multi Groups: Supervised Learning on Small Net-
Agent Based Simulation (MABS) of Finan- works,” 2017. https://aaai.org/ocs/index.
cial Transactions for Anti Money Launder- php/WS/AAAIW17/paper/view/15101 (accessed:
ing (AML),” pp. 25–32, 10 2012. http: 29.09.2017).
//www.diva-portal.org/smash/get/diva2:
[16] N.-A. Le-Khac, S. Markos, and M.-T. Kechadi,
834702/FULLTEXT01.pdf (accessed: 29.09.2017).
“A Heuristics Approach for Fast Detecting
Suspicious Money Laundering Cases in an
[9] N. LeKhac and M.-T. Kechadi, “Toward a
Investment Bank,” International Journal of
new cloud-based approach to preserve the pri-
Computer, Electrical, Automation, Control and
vacy for detecting suspicious cases of money
Information Engineering, vol. 3, no. 12, 2009.
laundering in an investment bank,” Interna-
http://waset.org/publications/12189/a-
tional Conference on Computational Science,
heuristics-approach-for-fast-detecting-
2014. http://researchrepository.ucd.
suspicious-money-laundering-cases-in-an-
ie/bitstream/handle/10197/6553/insight_
investment-bank (accessed: 29.09.2017).
publication.pdf?sequence=1 (accessed:
29.09.2017). [17] N. A. L. Khac, S. Markos, and M.-T. Kechadi, “A
Data Mining-Based Solution for Detecting Suspi-
[10] A. K. A. A. El-Din and N. E. Khamesy, “Data cious Money Laundering Cases in an Investment
Mining Techniques for Anti Money Laundering,” Bank,” pp. 235–240, 2010.
International Journal of Advanced Research in
Science, Engineering and Technology, vol. 146, [18] M. P. S.-B. Almeida, “Classification for fraud
2016. detection with social network analysis,” Masters
Degree Dissertation, Engenharia Informatica e de
[11] A. S. Larik and S. Haider, “Clustering based Computadores, 2009.
anomalous transaction reporting,” Procedia
Computer Science, vol. 3, pp. 606 – 610, 2011. [19] A. Samantha Maitland Irwin, K.-K. Ray-
http://www.sciencedirect.com/science/ mond Choo, and L. Liu, “An analysis of money
article/pii/S187705091000476X (accessed: laundering and terrorism financing typologies,”
29.09.2017). Journal of Money Laundering Control, vol. 15,
no. 1, pp. 85–111, 2011. http://search.
[12] K. D. Rohit and D. B. Patel, “Review On De- ror.unisa.edu.au/media/researcharchive/
tection of Suspicious Transaction In Anti-Money open/9915914120301831/53109886220001831
Laundering Using Data Mining Framework,” In- (accessed: 29.09.2017).
ternational Journal for Innovative Research in
[20] I. Kozlov, “About the stages of the money laun-
Science & Technology, vol. 1, 2015. http://www.
dering process,” Finance and credit, vol. 31,
ijirst.org/articles/IJIRSTV1I8043.pdf (ac-
no. 278, pp. 76–79, 2007.
cessed: 29.09.2017).
[21] Guide to U.S. Anti-Money Laundering Require-
[13] K. Manjunath, “Data Mining Techniques ments. https://www.protiviti.com/sites/
for Anti Money Laundering,” International default/files/united_states/insights/
Journal of Advanced Research in Science guide-to-us-aml-requirements-6thedition-
Engineering and Technology, vol. 2, 2015. protiviti_0.pdf (accessed: 29.09.2017).
http://www.ijarset.com/upload/2015/
august/1_IJARSET_manjunath.pdf (accessed: [22] Australian Registered AML/CTF Rules.
29.09.2017). http://www.austrac.gov.au/businesses/
7
legislation/registered-amlctf-rules (ac- https://pdfs.semanticscholar.org/5c31/
cessed: 29.09.2017). e690df172d631b933dd60a81a168f928d02d.pdf
(accessed: 29.09.2017).
[23] Financial Industry Regulatory Authority.
http://www.finra.org/industry/aml (ac- [31] C. Wang, K. Viswanathan, L. Choudur, V. Tal-
cessed: 29.09.2017). war, W. Satterfield, and K. Schwan, “Sta-
tistical Techniques for Online Anomaly Detec-
[24] A. S. M. Irwin, K.-K. R. Choo, and
tion in Data Centers,” IFIP/IEEE International
L. Liu, “Modelling of Money Launder-
Symposium on Integrated Network Management,
ing and Terrorism Financing Typologies,”
2011. http://www.hpl.hp.com/techreports/
Journal of Money Laundering Control,
2011/HPL-2011-8.pdf (accessed: 29.09.2017).
vol. 15, no. 3, pp. 316–335, 2012. https:
//www.researchgate.net/publication/ [32] V. Chandola, A. Banerjee, and V. Kumar,
263332155_Modelling_of_Money_Laundering_ “Anomaly Detection: A Survey,” ACM Com-
and_Terrorism_Financing_Typologies (ac- puting Surveys, vol. 41, no. 15, 2009. http:
cessed: 29.09.2017). //cucis.ece.northwestern.edu/projects/
DMS/publications/AnomalyDetection.pdf
[25] L. S. Bershtein and A. A. Tselykh, “A Clique-
(accessed: 29.09.2017).
based Method for Mining Fuzzy Graph Patterns
in Anti-money Laundering Systems,” pp. 384– [33] S. Omar, A. Ngadi, and H. H. Jebur, “Ma-
387, 2013. http://doi.acm.org/10.1145/ chine Learning Techniques for Anomaly Detec-
2523514.2523568 (accessed: 29.09.2017). tion: An Overview,” International Journal of
[26] N. Heidarinia, A. Harounabadi, and Computer Applications, vol. 79, no. 2, pp. 33–
M. Sadeghzadeh, “An Intelligent Anti-Money 41, 2013. http://research.ijcaonline.org/
Laundering Method for Detecting Risky Users volume79/number2/pxc3891478.pdf (accessed:
in the Banking Systems,” International Jour- 29.09.2017).
nal of Computer Applications, vol. 97, no. 22, [34] C. Phua, V. Leea, K. Smith, and R. Gayler, “A
2014. http://research.ijcaonline.org/ Comprehensive Survey of Data Mining-Based
volume97/number22/pxc3897780.pdf (accessed: Fraud Detection Research,” 2010. https:
29.09.2017). //www.google.ru/url?sa=t&rct=j&q=&esrc=
[27] R. Drezewski, J. Sepielak, and W. Filip- s&source=web&cd=10&ved=0ahUKEwihoJ-
kowski, “The Application of Social Network ijYHRAhUDiiwKHZyGBHAQFghbMAk&url=
Analysis Algorithms in a System Support- https%3A%2F%2Fsites.google.com%2Fsite%
ing Money Laundering Detection,” Inf. Sci., 2Fcliftonphua%2Ffraud-detection-
vol. 295, no. C, pp. 18–32, 2015. https: survey.pdf&usg=AFQjCNEWy-HRGI-
//www.researchgate.net/publication/ yrgXlOcY5lOkDmn9V5w&sig2=5bIw1d1Z5wtoyn_
267632670_The_application_of_social_ XCNAykA&cad=rjt (accessed: 29.09.2017).
network_analysis_algorithms_in_a_system_
[35] A. D. Pozzolo, O. Caelen, Y.-A. L. Borgne, S. Wa-
supporting_money_laundering_detection
terschoot, and G. Bontempi, “Learned lessons in
(accessed: 29.09.2017).
credit card fraud detection from a practitioner
[28] Z. Li, H. Xiong, and Y. Liu, “Detecting Black- perspective,” Expert Systems with Applications,
holes and Volcanoes in Directed Networks,” vol. 41, pp. 4915–4928, 2014.
CoRR, 2010. https://arxiv.org/pdf/1005.
[36] R. Drezewski, G. Dziuban, ukasz Hernik, and
2179.pdf (accessed: 29.09.2017).
M. Paczek, “Comparison of Data Mining Tech-
[29] T. E. Senator, H. G. Goldberg, and J. W. et al., niques for Money Laundering Detection System,”
“Financial Crimes Enforcement Network AI Sys- Conference: 2015 International Conference on
tem (FAIS) Identifying Potential Money Launder- Science in Information Technology, 2015.
ing from Reports of Large Cash Transactions,” AI
Magazine, vol. 16, no. 4, pp. 580–585, 1995. [37] K. Michalak and J. Korczak, “Graph Min-
ing Approach to Suspicious Transaction Detec-
[30] S.-N. Wang and J.-G. Yang, “A Money Laun- tion,” Proceedings of the Federated Conference
dering Risk Evaluation Method Based on on Computer Science and Information Systems,
Decision Tree,” Machine Learning and Cyber- 2011. https://fedcsis.org/proceedings/
netics, 2007 International Conference on, 2007. 2011/pliks/218.pdf (accessed: 29.09.2017).
8
[38] P. Berkhin, “Survey of Clustering Data Mining [47] B. van Liebergen, “Machine Learning: A Revo-
Techniques,” http://www.cc.gatech.edu/ lution in Risk Management and Compliance?,”
~ isbell/reading/papers/berkhin02survey. The Capco Institute Journal of Financial Trans-
pdf (accessed: 29.09.2017). formation, vol. 4, 2017. https://www.iif.com/
system/files/32370132_van_liebergen_-
[39] R. Winkler, F. Klawonn, and R. Kruse, “Prob- _machine_learning_in_compliance_risk_
lems of Fuzzy c-Means Clustering and Similar management.pdf (accessed: 29.09.2017).
Algorithms with High Dimensional Data Sets,”
2010. [48] E. A. Lopez-Rojas and S. Axelsson, “Money
Laundering Detection using Synthetic Data,” The
[40] P. Kanhere and H. K. Khanuja, “A Sur- 27th annual workshop of the Swedish Artificial
vey on Outlier Detection in Financial Intelligence Society (SAIS), 2012. http://www.
Transactions,” International Journal of ep.liu.se/ecp/071/005/ecp12071005.pdf (ac-
Computer Applications, vol. 108, no. 17, cessed: 29.09.2017).
2014. http://research.ijcaonline.org/
volume108/number17/pxc3900502.pdf (ac- [49] Australian Transaction Reports and Analysis
cessed: 29.09.2017). Centre. http://www.austrac.gov.au (accessed:
29.09.2017).
[41] J. M. Keller, M. R. Gray, and J. James
A. Givens, “A Fuzzy K-Nearest Neighbor Algo- [50] D. Savage, Q. Wang, P. Chou, X. Zhang,
rithm,” IEEE Transactions on Systems, Man, and X. Yu, “Detection of money laundering
and Cybernetics, vol. SMC-15, no. 4, pp. 580–585, groups using supervised learning in networks,”
1985. https://myteks.files.wordpress.com/ 2016. https://arxiv.org/pdf/1608.00708.
2012/03/fuzzyknn.pdf (accessed: 29.09.2017). pdf (accessed: 29.09.2017).
[42] J. Kingdon, “AI Fights Money Laun- [51] J. Yang, J. McAuley, and J. Leskovec, “Com-
dering,” IEEE Intelligent Systems, munity Detection in Networks with Node At-
vol. 19, pp. 87–89, may 2004. https: tributes,” 2014. https://arxiv.org/pdf/1401.
//pdfs.semanticscholar.org/0380/ 7267v1.pdf (accessed: 29.09.2017).
c04339f606413c71b657fab8ee3e4556c046.pdf
(accessed: 29.09.2017).
[43] L. Akoglu, H. Tong, and D. Koutra, “Graph
Based Anomaly Detection and Description: A
Survey,” Data Min. Knowl. Discov., vol. 29,
pp. 626–688, May 2015. https://arxiv.org/
pdf/1404.4679.pdf (accessed: 29.09.2017).
[44] M. Pironet, C. Antunes, P. Moura, and J. Gomes,
“Classification for Fraud Detection with So-
cial Network,” https://fenix.tecnico.
ulisboa.pt/downloadFile/395138965337/
ResumoAlargadoFinalEntregue.pdf (accessed:
29.09.2017).
[45] G. Hamerly and C. Elkan, “Alternatives
to the K-means Algorithm That Find
Better Clusterings,” pp. 600–607, 2002.
http://people.csail.mit.edu/tieu/
notebook/kmeans/15_p600-hamerly.pdf (ac-
cessed: 29.09.2017).
[46] E. Frank and I. H. Witten, “Gen-
erating Accurate Rule Sets Without
Global Optimization,” pp. 144–151, 1998.
https://pdfs.semanticscholar.org/3998/
12d46345ad7d93f5510b1bbda30948e7a65c.pdf
(accessed: 29.09.2017).
9