Modernizing Check Fraud Detection with Machine Learning

by

Lydia M. Rose

A Capstone Project Submitted to the Faculty of

Utica College

December 2018

in Partial Fulfillment of the Requirements for the Degree of

Master of Science in
Financial Crime and Compliance Management




ProQuest Number: 13421455




All rights reserved

INFORMATION TO ALL USERS
The quality of this reproduction is dependent upon the quality of the copy submitted.

In the unlikely event that the author did not send a complete manuscript
and there are missing pages, these will be noted. Also, if material had to be removed,
a note will indicate the deletion.






ProQuest 13421455

Published by ProQuest LLC (2018 ). Copyright of the Dissertation is held by the Author.


All rights reserved.
This work is protected against unauthorized copying under Title 17, United States Code
Microform Edition © ProQuest LLC.


ProQuest LLC.
789 East Eisenhower Parkway
P.O. Box 1346
Ann Arbor, MI 48106 - 1346
© Copyright 2018 by Lydia M. Rose

All Rights Reserved

ii
 Abstract

Even as electronic payments and virtual currencies become more popular, checks are still the

nearly ubiquitous form of payment for many situations in the United States such as payroll,

purchasing a vehicle, paying rent, and hiring a contractor. Fraud has always plagued this form of

payment, and this research aimed to capture the scope of this 15th century problem in the 21st

century. Today, counterfeit checks originating from overseas are the scourge of online dating

sites, classifieds forums, and mailboxes throughout the country. Additional frauds including

alteration, theft, and check kiting also exploit checks. Check fraud is causing hundreds of

millions in estimated losses to both financial institutions and consumers annually, and the

problem is growing. Fraud investigators and financial institutions must be better educated and

armed to successfully combat it. This research study collected information on the history of

checks, forms of check fraud, victimization, and methods for check fraud prevention and

detection. Check fraud is not only a financial issue, but also a social one. Uneducated and

otherwise vulnerable consumers are particularly targeted by scammers exploiting this form of

fraud. Racial minorities, elderly, mentally ill, and those living in poverty are disproportionately

affected by fraud victimization. Financial institutions struggle to strike a balance between

educating customers, complying with regulations, and tailoring alerts that are both valuable and

fast. Applications of artificial intelligence including machine learning and computer vision have

many recent advancements, but financial institution anti-fraud measures have not kept pace. This

research concludes that the onus rests on financial institutions to take a modern approach to

check fraud, incorporating machine learning into real-time reviews, to adequately protect

victims. Keywords: Financial Crime and Compliance Management, Dr. Kyung-Seok Choo,

payment fraud, behavioral alerts, banks, credit unions, false positive.

iii
 Acknowledgments

My first and greatest acknowledgement must be of my husband, Michael. His steadfast support

of my endeavors (including spending the first two years of our marriage completing a master’s

program) can only be repaid with a lifetime of my love. I hope to support you as loyally in

whatever you wish to do. Thank you, Michael, and I cannot wait to be your wife again! Second, I

would like to thank the faculty and staff of Utica College for leading students on an academically

rigorous course of study. Completing the Financial Crime and Compliance Management program

has prepared me for the rest of my career and served to further ignite my passions for both justice

and learning. I am also deeply appreciative for the support of my friends, family, and colleagues

who have collectively buoyed me on this journey. I doubt my sanity would be as nearly intact

without you all and I appreciate you sticking by me. I owe a particular thank you to my manager,

Mike McWhirt, whose mentorship and guidance—both on this paper and always—are endlessly

respected. Finally, to the management of Credit Union of Colorado, thank you for the support

you give to employees to follow their dreams. I assure you all that your investment in me is

cherished and will not be wasted. Thank you.

iv
 Table of Contents

List of Illustrative Materials ............................................................................................vi

Introduction ..................................................................................................................... 1
Literature Review ............................................................................................................ 8
A brief history of checks ............................................................................................ 8
Methods of check fraud............................................................................................ 11
Victim information .................................................................................................. 15
Evaluation of current anti-fraud measures ................................................................ 20
Machine learning ..................................................................................................... 24
Summary ................................................................................................................. 34
Discussion of the Findings ............................................................................................. 35
Major findings ......................................................................................................... 36
Comparison of the findings with existing studies ..................................................... 37
Limitations of the study ........................................................................................... 40
Recommendations ......................................................................................................... 42
Recommendations for future research ...................................................................... 42
Recommendations for combatting check fraud ......................................................... 44
Conclusion .................................................................................................................... 48
References ..................................................................................................................... 50

v
 List of Illustrative Materials

Figure 1 – Process to commit fraud................................................................................ 14

Figure 2 – Fraud alert process ........................................................................................ 26
Figure 3 – Altered money order ..................................................................................... 31
Figure 4 – Example postal money order ......................................................................... 32

vi
 Introduction

Check fraud is still a modern-day problem and necessitates a modern-day solution.

Ancient merchants required a secure method of transferring funds from one person to another

over a great distance. Beginning in the 1200s, they used an instrument called a bill of exchange,

which was later replaced by the widespread use of checks. The first check fraud was reported in

the fifteenth century (Quinn & Roberds, 2008, p. 3). Today, though the use of checks has

decreased year over year since 1999, the decline has slowed. The dollar value of commercial

checks clearing through the federal reserve system actually increased between 2016 and 2017

(Board of Governors of the Federal Reserve System, 2018b). Even with the arrival of Automatic

Clearing House (“ACH”) and virtual currency payment systems, the need for paper checks has

not been completely eradicated.

Check fraud has resurged in recent years. Both the number of overall consumer reports of

check fraud victimization and the proportion of check fraud events compared to all fraud types

increased between 2015 and 2017 (Federal Trade Commission, 2018). In a survey of 138 banks,

the American Bankers Association found that check fraud was responsible for 35% of overall

fraud losses to banks in 2017. Bank respondents also reported an increase in check fraud

compared to 2008. This shift was attributed to the relative ease of committing check fraud while

card fraud has become more difficult with the recent implementation of chip cards (“EMV”)

(American Bankers Association, 2018). A coordinated effort by law enforcement and private

partners to concentrate on check fraud, through consumer education and enforcement efforts was

made in 2009. A decline in consumer complaints to the Federal Trade Commission (“FTC”)

Consumer Sentinel was observed following this effort but check fraud complaints have begun to

1
grow again in the last four years (Baker, 2018). Thus, it is highly likely that investigation and

implementation of modern check fraud detection techniques is a worthwhile undertaking.

The purpose of this research was to evaluate the current status of check fraud and propose

a modern solution. Who is involved in check fraud and how does it take place? How could

machine learning be used to combat check fraud? Is it superior to—or could it supplement—

existent solutions such as behavioral alerts and consortiums?

In one type of check fraud, victims are scammed into depositing (or allowing a fraudster

to deposit) fraudulent checks and money orders. According to a study by the Better Business

Bureau (“BBB”), checks are used in many different types of scams. For instance, the victim may

receive a check for more than expected for the sale of an item on an online classifieds site, with

instructions to purchase supplies for a work from home job, or to cover taxes on a supposed

sweepstakes prize (Baker, 2018). The victim is then directed to send a portion of the funds to the

perpetrator as a form of advance fee fraud. Advance fee fraud is characterized by an upfront

payment by the victim to the fraudster with the expectation that the payment will lead to future

profit or benefit. After the fee is advanced, no profit or benefit is ever received (Akinladejo,

2007, p. 321).

Several factors have conspired to make check scams especially damaging when

successful. Victims are psychologically primed to respond to scams due to perceptions of

scarcity (possibility of losing out on a perceived opportunity) and reciprocation (complying with

the requests of someone to whom one feels a sense of indebtedness) (Archer, 2017, p. 35).

Unwittingly, the consumer’s exposure to this form of fraud increased as a result of the

Availability of Funds and Collection of Checks (“Regulation CC”). Regulation CC requires that

financial institutions make funds available for certain “low risk” items such as treasury and
2
cashier’s checks on the business day following the deposit while all other checks must have a

portion made available by the second business day (Board of Governors of the Federal Reserve

System, 2017). Many financial institutions expedite funds availability ahead of the schedule

required by regulation as a courtesy to their customers. The victim often advances the fee as

directed by the scammer from funds made available from the check deposit, mistakenly believing

the availability of the funds implies validation of the check.

Once the deposit is found to be fraudulent, the entirety of the balance is owed back to the

depository financial institution. Under their service agreements, most financial institutions deny

responsibility for validation of checks and place the liability for fraudulent deposits on the

customer who allowed the activity. For example, in Bank of America’s deposit agreement and

disclosures, the bank informs customers that it does not verify if checks, money orders, cashier’s

checks, or similar items are authentic and valid at the time of negotiation. Further, if an item is

later discovered to be fraudulent, counterfeit, or invalid for any reason, the customer’s account

can be charged for the full amount (Bank of America, 2018). Losses to the victim customer are

not covered by federal deposit insurance (Federal Deposit Insurance Corporation, 2018). If the

customer has no ability to repay, the balance is written off as a loss to the financial institution

and is assigned to collections. The collection process can result in civil judgements and

negatively affects the victim’s credit history.

Vulnerability to check fraud can be identified through several risk factors. A survey by

the AARP characterized victims of advance fee frauds as single, less than college educated,

earning less than $50,000 annually, and less aware of consumer protection information (Pak &

Shadel, 2011, p. 5). Individuals who recently suffered a serious loss like a death or divorce are at

higher risk of falling victim to a scam (Federal Trade Commission, 2013). In its annual report on
3
fraud, the Federal Bureau of Investigation reported that the majority of complaints it received

through its Internet Crime Complaint Center (“IC3”) were from victims over 60, who also

experienced the highest total losses in all types of fraud (Federal Bureau of Investigation, 2017).

Beyond likelihood for victimization, some consumers—particularly those that are disabled,

elderly, or impoverished—are disproportionately harmed by check fraud.

According to a survey of 31,980 victims, the median loss in a counterfeit check scam was

$1,008 (Federal Trade Commission, 2018). Victims aged 80 or older reported they lost at least

twice as much as those between 20 and 69, and 1.5 times as much as those between 70 and 79

(Federal Trade Commission, 2018). While most consumers reported they would use savings if an

unexpected expense arose, for individuals living on the edge of homelessness or bankruptcy, a

$1,000 loss may spell ruin. The median balance in savings for households earning under $25,000

a year is $670. The median savings account balance for African American and Hispanic

households is $1,500 (Board of Governors of the Federal Reserve System, 2018a). The

consequences of check fraud victimization may seem trivial in comparison to thousands or

millions lost in investment schemes and other scams. However, beyond financial strain, fraud

victims experience other serious consequences including feelings of shame, depression, and even

suicidal ideation (Ganzini, McFarland, & Bloom, 1990, pp. 61-62). Check fraud, while it can

happen to anyone, impacts the vulnerable most severely. Fortunately, it is also one of the most

consistent scams to profile because all check fraud begins with a single predicate act: the

negotiation of the item at a financial institution.

Financial institutions are in the best position to combat the avoidable loss caused by

check fraud, and it is in their own interest. In 2016, check fraud cost banks approximately $789

million, an increase of 25% since 2014 (Baker, 2018). Customers are dissatisfied by the
4
experience of falling for a scam and losing money. They lose trust in their bank’s security

practices and have incentive to terminate the relationship (Telang & Somanchi, 2016). Banks

already perform substantial fraud prevention, having stopped around $9 out of every $10 in

attempted deposit account fraud in 2016 (American Bankers Association, 2018). Much of this

success can be attributed to the training and experience of front-line staff and bank efforts to

educate the public. However, almost 75% of the time, the customer withdraws the money from a

fraudulent check at a point of sale, not at a teller line (Baker, 2018). This suggests further effort

must be made to detect fraudulent items at the time of deposit—before any funds are spent.

One potential solution is machine learning. Machine learning is an application of

artificial intelligence (“AI”) that allows computers to learn and improve from previous

experiences without manual programming (Jean-Philippe, 2018, p.1). Machine learning can be

utilized for advanced decision making and pattern recognition of text, speech, and images. Some

forms of machine learning attempt to bring computers as close as possible to the point of

mirroring human reasoning when a pre-determined answer cannot be applied to the data

(Lidestri, 2018, pp. 5-6). Machine learning can still be designed to maintain logical neural

pathways while remaining free from preconceived biases and flaws of human analysts. An

adjuacent field to machine learning is data mining. As explained by Dua & Du (2011), data

mining is the extraction of useful information from large amounts of data. The patterns,

relationships, and rules determined by data mining techniques can be used to predict behavior

from patterns discovered in the data (p. 5). Hitherto unrecognized information within large sets

of data, such as fraud transactions, can be discovered and understood by applying AI, machine

learning, and data mining.

5
 The financial industry is already beginning to use AI and machine learning. Machine

learning technologies are being applied to predicting credit default (Moula, Guotai, & Abedin,

2017) and events in the stock market (Chatzis, Siakoulis, Petropoulos, Stavroulakis, &

Vlachogiannakis, 2018, p. 353). The Securities and Exchange Commission (“SEC”) recently

began using a machine learning approach to behavioral predictions and risk assessment.

Researchers built topic and sentiment models to examine text in enforcement actions and

examination results, which could be used to identify future fraud and misconduct (Kutler, 2017).

The Financial Industry Regulatory Authority (“FINRA”) reported success with a machine

learning feedback loop between human analysts and a program based on current trends and

patterns, allowing more nimble fraud detection (Financial Industry Regulatory Authority, 2018).

Machine learning algorithms have even been developed to create “computer vision” which

senses variation in light and object shapes in images and video, mimicking the sensorial areas of

the neocortex in the human brain (Sohangir, Wang, Pomeranets, & Khoshgoftaar, 2018, p. 2).

This paper will theorize it could be beneficial to use machine learning to “read” the check with a

more human-like eye and recognize red flags, alterations, or signs of a counterfeit instrument.

Current solutions for check fraud prevention and detection, other than bank efforts

mentioned above, are two-fold. One method fraud investigators at financial institutions use is

software that generates alerts based on transaction activity. As described by Julisch (2010), the

two most common classifications of alerts are knowledge-based and behavior-based.

Knowledge-based alerts use knowledge of past frauds to detect future instances of the same

pattern of fraud. This is a simple form of machine learning. Behavior-based alerts detect

deviations from parameters identified as “normal” or non-fraudulent activity. Most fraud

detection systems use a combination of both knowledge-based and behavior-based software

6
alerts (p.5). Alerts may generate in as the transaction occurs (real-time) or in batches over some

period of time such as a business day, depending on the financial institution’s practices. One of

the issues with this approach to check fraud detection is false positive and false negative alerts. A

false positive is an alert mistakenly triggered when the condition is not present, while a false

negative is when an alert fails to trigger despite the condition actually being present (Rouse,

2014). Another problem is a delay between the time of deposit, the triggering of the alert, and the

review of the alert by an analyst. At any point, the participant may have already withdrawn the

funds and the detection of the fraud will be too late.

The second solution commonly used by financial institutions is check fraud consortiums.

One example is a company called Early Warning, which has developed a bank-to-bank network

of 2,300 financial institutions including five of the largest banks in the U.S. Members of the

network can query the database of collaborative, shared intelligence on available balances,

account status, and more (Early Warning, 2015). The member financial institutions consort with

each other, making identification of potential problem items far easier. As with the above

software alerts, consortiums create many false positive alerts. Both software alerts and

consortiums cannot be discounted for their utility, but financial institutions could still do more to

fight check fraud more effectively.

Checks are a centuries-old method of payment. Many new forms of payment have been

developed, and yet the check has persisted. Scammers have taken advantage of modern

technology like Photoshop to make better fraudulent checks, but there is still room to innovate in

check fraud prevention. There are many possible solutions offered by present-day technology.

Whether drawn in by the promise of a new job, online romance, or a lottery prize, at-risk

7
consumers experience dire consequences from negotiating fraudulent checks. Protecting victims

of check fraud is a moral—and financial—imperative.

Literature Review

The purpose of this research was to evaluate the current status of check fraud and propose

a modern solution. In light of this goal, one must understand how check fraud has evolved, the

relevant laws and regulations, and the capabilities of cutting-edge artificial intelligence

applications in this area. This portion of the paper will identify current industry efforts by

financial institutions to prevent and detect check fraud and explore the possibility of

technological enhancements to that end.

A brief history of checks

A check is a seemingly simplistic and archaic form of payment that facilitates the transfer

of funds from one party to another, with the help of one or more banks. The proper definition of

a check is as follows:

A check is a written order to a bank by a depositor at that bank (known as the drawer or

maker). The order instructs a bank to pay a third party (known as the payee) a certain sum

of money from the depositor’s account when the check is presented to the bank (Quinn &

Roberds, 2008).

Routing and account number information along the bottom of the check indicates which account

at which depository bank should be debited to pay the item. A check is, at its most basic, a piece

of paper with a few special words and numbers inscribed upon it; yet, this form of payment has

been in use for at least 600 years. It is important to understand the history of the check and its

true complexity, despite its evident simplicity.

8
 Quinn & Roberds (2008) completed a historical survey of the check as a method of

payment from medieval times to present for the Federal Reserve Bank of Atlanta. The word

‘check’ may have similar origins to the Arabic sakk, an instrument distributed in the Muslim

world in the tenth and eleventh centuries which allowed merchants to pay third parties from their

bank accounts via written order. Another instrument, called a bill of exchange, was also

commonly used by merchants in the same time period in Europe. A weakness of the bill of

exchange was that it only worked over distance as it relied on the passage of time: generally, the

length of time it took for the ship carrying goods to arrive to its destination. The check

superseded the bill of exchange in popularity by the sixteenth century in Europe with the

invention of negotiability, or the item being recognized as collectible after circulation. The

second revolutionary change in the same period was the practice of endorsement, which allowed

transfer of the note to an endorsee and created liability for the endorser should the item be

dishonored. However, the check did not truly rise to primacy in the United States until the 1920s

with the establishment of the Federal Reserve nationwide clearing system (pp. 1-23).

The nationwide clearing system allowed the exchange of checks across state lines without

the use of complicated and expensive correspondent relationships that were previously relied

upon. The Federal Reserve collects checks for depository financial institutions, which receive

check deposits drawn on other institutions. The Federal Reserve Bank facilitates crediting the

depository account and collecting from the paying bank. Presentment of the item, reconciliation

with the paying bank, and notification of the paying bank’s response may take as little as one

business day each (Board of Governors of the Federal Reserve System, 2016). Though credit

cards began being used commonly in the 1960s and direct deposit or ACH payments were used

9
nationwide by 1978, checks remained the majority non-cash monetary payment method through

2003 (Quinn & Roberds, 2008, p. 23).

Though the check has since been overtaken in popularity by the debit card and a

combination of electronic payments, it has not been retired from use. In 2017, 5.15 million

checks were cleared through the Federal Reserve, valuing over $8.4 billion. This was an increase

of 4.3% over 2016 and the average daily value in commercial checks collected has increased

every year since 2013 (Board of Governors of the Federal Reserve System, 2018b). The number

of per capita payments made each year with a check declined from 148.4 in 2000 to 58.4 in 2012

in the United States. In the same period, the share of the number of payments made by check also

declined by 73% (Silva, Ramalho, & Vieira, 2017, p. 585). Despite this, the check is still popular

with both businesses and individuals in the United States. Checks function as a miniature

contract between the maker of the item, the paying institution, the payee, and the depository

financial institution (Quinn & Roberds, 2008, p. 7). Certain aspects of checks enhance the

verification of the transaction, namely maker signature and payee endorsement. The lack of

information exchange required between maker and payee makes the check a convenient form of

payment for personal instances like buying a car, paying rent, or settling a dentist bill. Finally,

details contained on checks like addresses and memo lines allow businesses to easily account for

payroll and invoicing (Monga, 2014). These features do not currently have a substitute amongst

electronic forms of payment.

Some critical developments have allowed the check to endure. The first was the

mechanization of check processing through the introduction of magnetic ink character

recognition (“MICR”) technology in 1956. The MICR line, as it is referred to in banking, is the

numerical information printed at the bottom of the check. The special ink used to print this line is
10
magnetic and readable both by people and by check scanners used by most financial institutions

(Quinn & Roberds, 2008, pp. 19-20). Next, the Expedited Funds Availability Act, which was

implemented by Regulation CC in 1987, addressed funds availability requirements for deposits,

including checks. Certain deposits, such as treasurer’s checks and cashier’s checks were

identified by Congress to be low risk and must have at least $200 made available on the business

day following the deposit while all other checks must have at least $200 made available by two

business days following the deposit (Board of Governors of the Federal Reserve System, 2017).

Regulation CC allows for exceptions for new accounts, large deposits, repeatedly

overdrawn accounts, emergencies including power failures, and more as long as proper

disclosure is given to the customer. This update facilitated consumer access to proceeds from

checks, making them more convenient, while also mandating timeframes for the return of

uncollectable items (Board of Governors of the Federal Reserve System, 2017). Finally, the

Check Clearing for the 21st Century Act (“Check 21”) enacted in 2003 allowed the use of

electronic images and processing for the collection of checks, creating efficiency in a heretofore

paper-heavy process (Board of Governors of the Federal Reserve System, 2016). The processing

of paper checks now ceases as soon as the item is converted into an electronic image, making it

almost unrecognizable from its ancient origins.

Methods of check fraud

Fraud is defined as a criminal act where a person uses deceit and dishonesty to obtain an

unjust advantage over another person or entity (Akinladejo, 2007, p. 320). In frauds involving

payment, the perpetrator exploits the credulity of the victim and deceives them into initiating

payments to the fraudster’s benefit (Julisch, 2010). Check fraud is as old as the instrument itself.

The check was first recorded in Europe around the year 1400 and with it came fraud, abuse, and
11
controversy. Once the concepts of negotiability and endorsement were established, when a check

was “dishonored” or unpaid due to fraud or default, the funds could still be pursued back along

the chain of endorsements based on the reputations staked by the prior endorsees of the item

(Quinn & Roberds, 2008, pp. 3-6). Collection on modern dishonored checks functions similarly.

By endorsing a check when depositing it at their financial institution, the payee warrants to their

bank that the check is good. As a check is processed, it may pass through several hands before it

reaches the bank it is drawn on (the paying bank). Each party provides a warranty to the next

party in the processing chain and unpaid funds, for fraud or otherwise, are still collected on the

basis of those warranties (Quinn & Roberds, 2008, p. 7).

Checks are subject to more fraud than any other payment method (Association for

Financial Professionals, 2018). According to the Association of Certified Fraud Examiners

(“ACFE”), check fraud is particularly common because some items required to commit check

fraud (scanner, printer, and personal computer) are easy to obtain and the risk of being caught

and prosecuted is low (Association of Certified Fraud Examiners, 2017, p. 431). There are

several different methods of check fraud. The ACFE’s Fraud Examiners Manual categorizes the

major forms as counterfeiting, theft, alteration, paperhanging, and kiting (Association of

Certified Fraud Examiners, 2017, pp. 426-428).

Counterfeit checks are fake checks purporting to be from legitimate individuals, financial

institutions, and businesses. Fraudsters attempt to copy legitimate formats of these instruments

and may use real routing and account numbers but the manufactured items are not authorized by

the true account holder (Salinger, 2005, p. 77). Fraudsters may obtain a legitimate check or

money order from the mail, for example, and then scan and Photoshop the original, editing it to

suit their needs before printing it on check stock for distribution (Baker, 2018). Counterfeit
12
checks attempt to replicate a template for a real check or money order while issuers of legitimate

items strive to create security features to thwart counterfeiters. Thus, an arms race continues

between the investigators and culprits of check fraud.

Theft and alteration are committed through use of legitimate checks rather than creating

original fictitious items. These two categories of check fraud often go hand in hand. A thief may

steal and forge blank checks, capture checks in transit, or take over an account and write checks

with the stolen information. Alteration occurs when the fraudster changes information on the

check (e.g. payee, amount) without the maker’s knowledge or permission before the item is

negotiated (Association of Certified Fraud Examiners, 2017, p. 427). The criminal utilizes a

chemical “wash” or other means to erase the targeted information and then adds information by

handwriting, typewriter, check imprinter, or laser printer to the check (Check Fraud Working

Group, 1999; Salinger, 2005, p. 77). Alteration and theft were likely forms of fraud that plagued

even the earliest forms of checks and monetary instruments, the bill of exchange and sakk.

The final two—paperhanging and kiting—are forms of check fraud wherein the customer

is not the victim, but the suspect. Paperhangers present knowingly bad checks (e.g. drawn on a

closed account) for goods or services at merchants and for cash at banks. They repeat the scam,

“hanging paper” or passing these checks and taking advantage of lax check hold and

identification review policies (Association of Certified Fraud Examiners, 2017, pp. 427-428).

Finally, kiting is a form of check fraud conducted by the customer themselves between one or

more of their own accounts. Check kiting takes advantage of what is known as check float. Float

is the delay between the time a check is deposited at the payee’s bank and the time the check is

paid by the paying bank (Quinn & Roberds, 2008, p. 21). A check kiter writes checks back and

forth from multiple accounts, but no funds actually exist. Until caught, this scheme can continue
13
indefinitely with one check written against insufficient funds covered by the next check deposit,

resembling a shell game (Association of Certified Fraud Examiners, 2017, p. 428; Salinger,

2005, p. 158).

Of the many varied forms of check fraud, counterfeit checks are the most often associated

with scams and are the form of check fraud most likely to cause a consumer loss. Most

counterfeit checks circulated in the United States originate from overseas (Baker, 2018). The

victim has little genuine identifying information regarding the actual counterfeiter in many cases

as the interaction occurs entirely online, by phone, or by mail (Association of Certified Fraud

Examiners, 2017, p. 426). Figure 1, below, illustrates the process of a check fraud facilitated by

an unknowing victim in a romance scam.

Figure 1. Process to commit fraud with a counterfeit check for romance scam. Reprinted from
Mobile Banking Fraud Trends: Fraudulent Checks Meet Remote Deposit Capture by Guardian
Analytics, 2016, Retrieved from https://guardiananalytics.com/wp-
content/uploads/2018/06/FraudUpdate_MobileRDCTrends_March2016.pdf. Copyright 2016 by
Guardian Analytics.

14
Online job, overpayment, and advance fee scams utilizing checks all follow similar models. An

estimated 400% increase was observed in losses between 2012 and 2014 due to fraudulent

checks deposited through Remote Deposit Capture (“RDC”) mobile banking services (Guardian

Analytics, 2016). The customer themselves either performs the check deposit or provides

account credentials that instigate the fraud. Conversely, in theft and alteration varieties of check

fraud, the account holder may have no awareness of the transactions until they review their

monthly bank statement. In cases of paperhanging and kiting, it is up to the financial institution

to detect and halt the activity, risking losses the longer the activity continues.

Victim information
The true victim of a check fraud may be difficult to identify. Liability for fraudulent

checks varies on the circumstances of the negotiation, timing of discovery, and type of fraud.

The customer is responsible for preventing and reporting forgeries, alterations, and other

unauthorized uses of their own checks. The financial institution is not required to perform

examination of checks outside the duty of ordinary care (Bank of America, 2018). The customer

is not typically held liable for unauthorized checks written from their accounts provided they

dispute the item in a timely manner. Financial institutions disclose deadlines for disputing

transactions in their agreements. The determination of whether the liability for the fraud lies on

the depository bank or the paying bank within that timeframe is governed by the Uniform

Commercial Code (“UCC”). The paying bank must either pay the item or return it by midnight

on the banking day following the day the item was presented for payment. Past this “midnight

deadline” the paying bank may still place the liability on the depository bank if the depository

bank breached its warranty in accepting the item. The paying bank may make a warranty claim

15
under the UCC for up to 3 years following the negotiation of the check (Smith, 2001).

Otherwise, the paying bank may be liable to reimburse the customer.

If their agreements allow for it, the depository financial institution may pass the liability

on to the customer that deposited the fraudulent item. If an item deposited to the customer’s

account has been paid by the bank on which it is drawn and that item is later returned with a

claim that the item was altered, forged, unauthorized, bears a forged or missing endorsement, or

should not have been paid, the customer’s account may be charged for the amount of that item

without prior notice (Bank of America, 2018). Customers who are victims of scams are not

covered by bank insurance policies, which cover the bank from losses related to fire, robbery,

and embezzlement, among others (Federal Deposit Insurance Corporation, 2018). Though liable,

the customer may refuse to pay, leaving the bank to attempt to gain restitution through criminal

or civil courts.

Likewise, the customer making the deposit is responsible if the deposit is returned unpaid

within the midnight deadline. A portion of the funds are made available to the customer by the

second business day following the deposit in most cases (Board of Governors of the Federal

Reserve System, 2017). In the majority of check scams, the victim sends funds made available

from the fraudulent check back to the scammer by wire transfer (Federal Trade Commission,

2018). The processing of the check through the Federal Reserve System may take up to 3 or

more business days. When the check is presented, the countdown for the midnight rule begins

(Smith, 2001). A check deposited on a Friday may not be returned until the following

Wednesday or even later, which is in compliance with UCC. This gap in time is where check

fraud losses to both customers and financial institutions occur.

16
 There are several sources to obtain data about check fraud victimization amongst

consumers. One is the FBI’s IC3 database. In 2017, IC3 received 301,580 complaints from

victims of Internet-facilitated crime. Of those, 16,368 reported being victims of advanced fee

scams and lost a collective $57.8 million, or about $3,535 per victim. Another 15,372 were

victims of confidence fraud or romance schemes ($13,750 loss per victim), 15,784 of

employment frauds (losing $2,463 on average), and 23,135 were victims of overpayment scams

($2,310 average loss) (Federal Bureau of Investigation, 2017). The data does not report how

many of these incidents involved check fraud. However, the FBI acknowledges that counterfeit

check schemes often target individuals that use the Internet for classifieds, romance,

sweepstakes, and employment (Federal Bureau of Investigation, n.d.).

Another source for check fraud victim data is the FTC’s Consumer Sentinel complaint

database. The FTC received reports from 2.7 million victims of fraud, identity theft, and other

schemes in 2017. Of those, 31,980 consumers reported they were victims of counterfeit check

scams. The number of reports of check fraud victimization from consumers to the FTC has

increased every year since 2015 (Federal Trade Commission, 2018). In fact, the number of

complaints received by the FTC Sentinel database and the IC3 more than doubled between 2014

and 2017 (Baker, 2018). Only 33% of all reporting consumers who were convinced to deposit a

fraudulent check as part of a scam said they took a loss. However, the average loss to consumers

in general from counterfeit checks was $1,008 in 2017, and the loss to a victim in the military

was $2,200 on average. These figures are significantly higher than the median loss of $429 per

consumer fraud incident overall (Federal Trade Commission, 2018). An estimated 36% of

businesses and 25% of households have recently been victims of white-collar crimes in contrast

with 8% for traditional property crimes and 1.1% for violent crime (Cliff & Wall-Parker, 2017)
17
 The National Consumers League, through their website Fraud.org, and the BBB’s Scam

Tracker are additional collectors of victim reports. Fake check scams were one of the top overall

scams reported to the National Consumers League, occurring both online and offline, and

complaints about fake checks were up 12% over the previous year (National Consumers League,

2018; Baker, 2018). Likewise, the BBB was contacted by 47,827 victims and found that fake

check/money order scams were a component of most other types of scams including employment

and sweepstakes. Check fraud victims had the third highest median dollar loss behind family

emergency and investment scams (Better Business Bureau, 2018).

Check fraud occurrences are likely vastly underreported. Only an estimated 29% of fraud

victims report to any sort of authority such as the Federal Trade Commission or BBB (Pak &

Shadel, 2011, p. 5) and less than 1 in 10 victims ever report to law enforcement (Baker, 2018).

Meanwhile, the United States Postal Inspection Service reported intercepting counterfeit checks

with a face value totaling $62 billion from the mail, and another 13,724 counterfeit postal money

orders totaling over $14 million in 2017 alone (Baker, 2018). The crimes not reported, and the

counterfeit items not cashed, serve to illustrate the immense problem of modern-day check fraud.

Based on reporting estimates, approximately 500,000 U.S. citizens were victims of check fraud

in 2017, losing a total of around $504 million (Baker, 2018).

As was mentioned in the introduction, certain demographics of individuals are more

likely to be victims of a check scam while certain ages, races, and income levels are

disproportionately harmed by the resulting financial loss. The highest rates of check fraud

occurred in victims in the age range of 20-29 (Baker, 2018). Those aged 20-29 reported losing

money in a scam 40% of the time, compared to only 18% of people aged 70 and older. However,

those over aged 70 had much higher median losses than any other age group (Federal Trade
18
Commission, 2018). African Americans and Hispanics were more likely to experience fraud than

other racial groups. Individuals making less than $20,000 a year and those who did not complete

high school were also more likely to become victims (Federal Trade Commission, 2013).

Surveys of consumer financial habits prove that the most vulnerable cannot sustain even

a $1,000 loss without being brought to the brink. Based on income, households making $25,000

annually have a median savings balance of $1,400, while those making less than $25,000 have

just $670 saved for an emergency. Conversely, households making $115,000 have a median of

$13,000 saved. By race, African American and Hispanic households have a median of $1,500

and $1,510, respectively, while Caucasian households have $9,700 (Board of Governors of the

Federal Reserve System, 2018a). These disparities reveal that even though anyone can be the

victim of check fraud, some are more adversely affected.

Even before the fraud occurs, victims’ lives often show signs of distress. Individuals who

engaged in risky purchasing practices, who had experienced a serious negative life event1 in the

past 2 years, or who indicated that they had more debt than they could handle were each more

likely to be victims of fraud than those who had not (Federal Trade Commission, 2013). Ganzini,

McFarland & Bloom (1990) performed a study comparing victims of violent crimes like rape,

robbery, and assault with victims of non-violent financial (“white-collar”) crimes of fraud,

forgery, and embezzlement. Their comparison was based on parameters including recovery rates

after victimization as well as the effects of the crimes on resultant psychiatric disorders. Both

victim groups were found to be more likely to require treatment for anxiety and major depressive

1
A serious negative life event was described in the survey to include a divorce, the death of a family member or

close friend, a serious injury or illness in the family, or the loss of a job (Federal Trade Commission, 2018).

19
disorder (up to and including suicidality) after the crime occurred. Of fraud victims, this was

particularly true for individuals who already had a history of mental illness or whom experienced

a decreased standard of living following the loss (pp. 55-62). Victims can be severely harmed by

check fraud.

Financial institutions also experience significant check fraud losses, either due to liability

under UCC rules or lack of repayment by customers (both victims and perpetrators). In a survey

of fraud and anti-money laundering professionals, 63% responded that scams are most frequently

funded with a fraudulent check and that more than a third of losses to their financial institution

were from customers who were scammed (Verafin, 2017a). Attempted fraud against deposit

accounts including checking accounts rose from $12.9 billion in 2014 to $19.1 billion in 2016.

Financial institution industry losses are estimated at $789 million annually as a result of check

fraud alone (American Bankers Association, 2018). 74% of organizations experienced check

fraud in 2017 (Association for Financial Professionals, 2018). Fraud losses to financial

institutions typically do not take into account expenses related to prevention, detection,

investigation, and prosecution of the fraud (Mazur, 2014, p. 11). According to the U.S.

Department of the Treasury, Financial Crimes Enforcement Network’s (“FinCEN”) Suspicious

Activity Report (“SAR”) stats, depository financial institutions filed 152,602 SARs in 2016 and

145,274 in 2017 regarding check fraud (Financial Crimes Enforcement Network, 2018). Check

fraud is a prolific issue.

Evaluation of current anti-fraud measures

Financial institutions are invested in preventing and detecting fraud both as a service to

their customers and to protect their reputation and assets (Julisch, 2010). Financial institutions

already employ many different measures to prevent and detect check fraud. Staff training is
20
important for identifying fraudulent checks. Frontline staff can recognize signs of alteration,

forgery, missing information, and red flags in customer statements and actions (Association of

Certified Fraud Examiners, 2017, pp. 433-434). Some banks offer positive pay to their

customers, which is a process for clearing checks only if they are verified against a list of

authorized items and payees provided by the customer (Association of Certified Fraud

Examiners, 2017, p. 204). When surveyed by the ABA, most banks reported using manual

reviews, transaction monitoring, call-back procedures, and out-of-pattern detection to detect

fraud (Mazur, 2014, p. 11). Transaction monitoring and out-of-pattern detection are most often

accomplished using alerts and consortiums.

One of the most common strategies today is transaction monitoring software2 that triggers

alerts for review by human analysts. The following activity exemplifies some patterns of check

fraud that may be detectable by fraud alert software: frequent deposits and withdrawals—

particularly round number or in the same amount, low average balance compared to high level of

sudden deposits, checks clearing out of sequence, and cash withdrawn immediately from

deposited checks (Association of Certified Fraud Examiners, 2017, pp. 434-435). Commercial

payment fraud detection systems use a combination of behavior-based and knowledge-based

detection methods. A knowledge-based alert would trigger if multiple transactions occur within a

given time period and aggregate above a given dollar threshold. Behavior-based alert techniques

focus on comparing transaction behavior to the expected value based on historical or global

norms (Julisch, 2010).

2
Prominent fraud detection software companies include Verafin, NICE Actimize, Fiserv, and Banker’s Toolbox.

Some financial institutions opt to develop proprietary transaction-monitoring software in-house.

21
 An important feature for adequate fraud alerts is the ability to detect cross-channel fraud,

or fraud that involves more than one type of product or service (Castanheiro, 2017). For instance,

in the type of fraud profiled in Figure 1 (above), fraud alert software may be able to detect what

is known as “account takeover” when the fraudster accesses the victim’s mobile banking account

to perform the check fraud. Characteristics of account takeover include an unusual digital

fingerprint and changing contact information (Verafin, 2017b). It is not economical nor possible

to prevent 100% of fraud (Julisch, 2010). When implemented, fraud detection through software

alerts can be tailored to acceptable risk thresholds set by the individual financial institution.

However, cross-channel considerations make risk-rating more effectual.

Fraud alerts are only as useful as they are practical. The practicality for financial

institutions relies on the cost of ownership, implementation, and operation of the software

compared with the fraud losses prevented as a result. Most financial institutions still experience

difficulty with excessive false positives and false negatives and a manual response to fraud (e.g.

blocking payments) following reception of the alert. Fraud alerts are also limited by the

software’s ability to adapt and evolve over time and the battle to anticipate rather than detect

fraud after the fact (Julisch, 2010). False identification of fraud and the inconvenience of

needlessly blocked payments or delayed release of funds impacts consumer sentiment.

Nonetheless, banks that display knowledge and competence regarding fraud create a feeling of

safety for their customers, improving key factors of customer relationship quality and loyalty.

These benefits extend beyond the reduction of fraud-induced operating costs and are shown to be

crucial for customers of all ages and income levels (Hoffmann & Birnbrich, 2012, p. 403).

Under Section 314b of the USA PATRIOT Act, financial institutions are permitted to

share information with each other regarding suspected money laundering, terrorist financing, and
22
other specified unlawful activity, including fraud (Department of the Treasury, 2009).

Participation in 314b information sharing is voluntary. Fraud prevention collaboratives called

consortiums have arisen based on a similar premise. Consortiums are a form of fraud prevention

based on information sharing between multiple member financial institutions. Consortium data

can make financial institution fraud detection efforts significantly better (Castanheiro, 2017).

One example is a data sharing program for closed accounts called ChexSystems. Participating

financial institutions report any checking account closed for fraud or other cause to the central

database and can query new applicants against the database prior to opening an account (Check

Fraud Working Group, 1999).

Early Warning is a consortium that integrates with check imaging systems (Early

Warning, 2015). At the time of deposit, member financial institutions can query the Early

Warning TrueChecks database, which is comprised of returned check data from financial

institutions across the United States. TrueChecks software delivers real-time responses on

counterfeit, non-sufficient funds, closed account, duplicate, and other returned items that

member financial institutions have reported (FLEX Credit Union Technology, n.d.). The goal of

consortiums is limiting the ability of fraudsters to reuse checks, identities, and accounts to

perpetrate fraud against multiple financial institutions (Castanheiro, 2017). Limitations of the

consortium model include data accuracy, membership size, and privacy regulations that limit

what information can be shared.

Finally, the criminalization of check fraud and prosecution of those responsible may deter

some from committing check fraud. Check fraud is primarily governed by state law in the United

States (Association of Certified Fraud Examiners, 2017, p. 445). Depending on the situation and

jurisdiction, perpetrators of check fraud may be prosecuted for felonies including bank fraud,
23
mail fraud, wire fraud, forgery, identity theft, or theft of a financial device among others

(Salinger, 2005, p. 74). For example, 18 U.S.C. § 1344 criminalizes defrauding financial

institutions as well as obtaining any of the funds owned by or under the custody or control of a

financial institution by means of false or fraudulent pretenses (Legal Information Institute, n.d.).

Because most American customers involved in negotiating counterfeit checks are merely

acting as “mules” of the funds to the fraudster, they are not often pursued with criminal charges

unless they can be found to be willfully and knowingly participating in the scheme (Manhattan

District Attorney's Office, 2018). However, federal law enforcement has experienced some

success in extraditing makers of fake checks from West Africa in the most egregious scams

(Baker, 2018). Nevertheless, in 2017 only 744 defendants were charged for theft by check and a

mere 57 for advance fee schemes by the United States Attorneys (Executive Office for United

States Attorneys, 2017). Arrests made for forgery and counterfeiting decreased by more than

50% from 2005-2014 (Cliff & Wall-Parker, 2017). Ultimately, check fraud and other white-

collar crimes are underprosecuted due to the difficulty of locating perpetrators and victims and

they often carry lighter sentences than violent crimes even when successfully prosecuted.

Machine learning

Artificial intelligence can be summarized as the ability of a computer system to perform

tasks normally reserved to humans such as seeing, recognizing speech, making decisions, and

translating languages (Jean-Philippe, 2018, p. 4). In the near future, use of artificial intelligence

will become imperative in detection of aberrant activity in applications spanning from fraud to

cyber security to physical security. Human analysis is not only fallible and slow, a workforce

shortage is expected based on predictions of future threat volumes in the fraud and security

industries (Lidestri, 2018, p. 2). The goal of AI is developing intelligent machines that think,
24
learn, see, and reason objectively, even surpassing human capabilities with methods of data

mining and machine learning in particular. Industry analysts recommend the use of more AI,

including machine learning, to prevent burnout for analysts otherwise performing tedious and

repetitive tasks. This reserves their energy for decisions requiring more nuanced judgment

(Lidestri, 2018, p. 12).

Machine learning is an application of AI that allows computers to recognize patterns and

intelligently make decisions through statistical modeling without requiring manual input from a

programmer. “Machine learning teaches a machine how to predict answers, which allows the

device to respond to situations it has not explicitly encountered before. The more examples the

invention sees in the form of a dataset the better the machine can learn and interpret” (Jean-

Philippe, 2018, p. 3).

The first step in machine learning is training. Depending on the availability of sample

data for training, as well as desired outcome, machine learning algorithms may be categorized as

supervised learning, unsupervised learning, or reinforcement learning (Jean-Philippe, 2018, p. 3).

In supervised learning, input and expected output data pairs are fed into a function and the

learning model is trained to predict output as efficiently as possible. Common sub-categories of

supervised learning algorithms include decision trees, artificial neural networks, and support

vector machines. In unsupervised learning, sample data is provided without expected output or

other labeling in order to allow the algorithm to determine which information is significant. The

most well-known groupings of unsupervised learning methods are hierarchical clustering and

self-organization maps (Dua & Du, 2011, pp. 7-8). In reinforcement learning, the model is

trained using unlabeled information and feedback is provided to it following its decision. The

model then uses the feedback to refine its actions over time (Jean-Philippe, 2018, p. 3).
25
 After the training stage, experts can set thresholds or controls for the machine learning

model to dictate certain responses based on criteria. Once deployed, machine learning can be

tuned based on performance with real data and improve over time with continued training. Much

like a person learns from past experience, reinforcement machine learning allows the algorithm

the ability to review old data and determine new patterns in light of new data given to the system

(Lidestri, 2018, p. 3). Machine learning may also be used for predictive modeling (e.g.

forecasting) and prescriptive frameworks (e.g. optimizing decisions) (Qiao & Beling, 2016).

Many financial institutions already utilize machine learning models in some form in their

fraud detection process, combining work by human analysts with model-driven software.

Knowledge-based alerts use fraud detection rules that were generated using supervised learning

techniques. Expected behavior for behavior-based alerts can be derived by use of link analysis,

clustering, and outlier detection techniques (Julisch, 2010). The process to examine and act on

the alerts is shown in Figure 2.

Figure 2. Contemporary process of transaction fraud alerts with human analysts and machine
learning models. Adapted from Solving the “False Positives” Problem in Fraud Prediction:
Automated Data Science at an Industrial Scale by R. Wedge, J. Kanter, K. Veeramachaneni, S.
Rubio, and S. Perez, 2017, Retrieved from http://arxiv-export-
lb.library.cornell.edu/pdf/1710.07709. Copyright 2017 by MIT.

26
A neural network machine learning model is already involved today when a financial

institution’s fraud software generates a risk score and combines it with rules created by experts

to provide alerts on potentially fraudulent transactions. Despite the use of machine learning,

human involvement in the process is expensive and false positives plague these alert systems

(Wedge, Kanter, Veeramachaneni, Rubio, & Perez, 2017). Fraudulent activity is missed if it falls

outside of parameters set, but thresholds must be narrow to avoid an overwhelming number of

alerts.

Wedge et al. (2017) studied the use of a technique called Deep Feature Synthesis to speed

up and improve the process of creating features for the machine learning model they used called

a random forest. The goal of the study was to address false positives in fraud prediction. An

estimated 4 out of 5 fraud alerts are not true fraud. Using an automated method, more than 100

behavioral patterns and 237 features based on historical data of the account were determined and

used to train the model. The model was 91.4% more accurate than the existing fraud alert

software used by the participant bank, BBVA. Because losses take place unless the fraud is

detected near instantaneously, machine learning was thought to require constant streaming of

data to be calculated. However, Wedge et al. (2017) found that aggregating transactional

information from the recent past as infrequently as every 35 days maintained predictive

accuracy.

In the false positive study, patterns and features were recognized by Deep Feature

Synthesis based on analysis of transactions along with the features that describe each transaction

(such as amount, time, location) and relationships (such as between transactions and accounts or

merchants and transactions). The algorithm synthesized potentially meaningful relationships

from patterns and features of the dataset it observed. An important decision the researchers made
27
was to select the training data to keep all transactions for a given card if it was included in the

training set in order to observe patterns of behavior. The machine learning algorithm returned

unique features for each account aggregated from all the past transactions including the mean

transaction amount, the average time between subsequent transactions, and the number of

distinct uses on unique days. This information was then useful to detect outlying activity, which

may be indicative of fraud (Wedge, Kanter, Veeramachaneni, Rubio, & Perez, 2017).

Another critical financial institution process adjacent to fraud detection is credit default

prediction. Matsatsinis (2002) posited that qualitative variables—including age, family status,

and years in profession—have an essential role in credit evaluation and must be included in

addition to quantitative measures—income, bank balance and credit score—for algorithms to

make appropriate credit decisions (p. 244). Wu, Hu, & Huang (2014) innovated further in

discovering that credit risk evaluations were more effective if performed in two stages. The first

stage involved feature selection and data pre-processing and the second stage utilized several

classification techniques to construct prediction models. The results were nine credit risk levels,

not just two classes (i.e. good credit or bad credit), which allow financial institutions to accept

applications within their risk appetite rather than simply approve or deny (pp. 1098-1101).

Moula, Guotai, & Abedin (2017) studied the application of a form of supervised machine

learning called a support vector machine to predict customers who likely may not repay their

loans if offered one. Previous researchers attempted to apply artificial neural networks for credit

default prediction, while this study aimed to improve on their results. The authors attempted to

create accurate and consistent decision-making using data sets from real world credit

applications. The data included facts from the customer’s application including financial, non-

financial, demographic, and social-graphic. Moula et al. used the information to train the model
28
to predict credit status of potential customers in a more precise and flexible manner than studies

done by previous scientists. The findings showed that the more variables collected during the

application process, the better the prediction ability of the model. Moreover, the study showed

that success measurement for default prediction, like fraud prediction, relies upon delineating as

accurately as possible between good and bad accounts (pp. 158-186). Customer financial status

modeling is an important facet to consider in all account relationships a financial institution may

extend—including services like checking, savings, debit cards, and check cashing.

Data mining may involve the application of statistical or machine learning techniques, or

both, to collect and analyze large amounts of data. It can be performed both by human analysts

and, to some extent, by artificial intelligence (Wu, Hu, & Huang, 2014). Westphal (2009)

explained that data mining for fraud detection requires investigators be well versed in technical

skills including databases and visualizations to effectively use their data. To be effective is to

understand complex meanings behind data connections and be able to interpret results from a

series of connected dots (p. 407). Financial institutions and businesses that market to them have

found it lucrative to perform data mining of the available customer information to better

understand, market to, and protect their customers (Salinger, 2005, p. 77).

According to Agarwal, Gupta, & Hussein (1997), computers have long had the ability to

“read” checks. During the analysis and recognition process, the computer identifies components

in the image on the basis of proximity and horizontal alignment of characters. Extraneous

information like background pictures must be filtered out as “noise.” Blocks or areas of the

check must be identified—payee, amount, maker, and bank routing and account information—all

while correcting for handwriting. These components are located based on rules and heuristics.

The computer recognizes the courtesy amount, for example, by looking for the dollar symbol ($).
29
The components are combined after this pre-processing and sent through neural network

recognition for verification (pp. 623-625).

Many abilities of extant check reading software have applicability to fraud detection. For

example, the computer can detect the difference between handwritten and machine printed text,

separate two handwritten characters that have been joined together, and determine fractions

based on differences in relative sizes of the characters (Agarwal, Gupta, & Hussein, 1997, pp.

635-637). As the research of Krizhevsky, Sutskever, & Hinton (2012) will show, these

capabilities may be built upon. With modern computer vision technology, instead of smoothing

and disregarding background noise, more components of the check could be included in the

analysis for fraud.

Computer vision is a sub-field within artificial intelligence, like machine learning

(Lidestri, 2018, p. 8). One of the human abilities AI seeks to mimic is visual perception.

Krizhevsky, Sutskever, & Hinton (2012) used a machine learning method called a deep

convolutional neural network to enhance computer object recognition. Their innovation over

previous researchers was to encourage the model to not “overfit,” meaning recognize the images

it was trained on well but do poorly with novel images. The researchers also used a training set

of data with 1.2 million images, which is much larger than the sets used in previous studies. Yet,

they were able to complete their training much faster than attempts by prior researchers, likely

due to their use of the deep convolutional neural network type. Minimal pre-processing of the

images—standardizing size and normalizing brightness—was found to be necessary to reduce

errors in the responses (pp. 1-8). Advances have recently been made with the same machine

learning method, deep convolutional neural networks, using the internal structure that exists

inside text documents for tagging, entity search, sentence modeling, and more (Sohangir, Wang,
30
Pomeranets, & Khoshgoftaar, 2018, p. 2). Computer vision relies in part on calculating similarity

of a given image to images the algorithm was trained with. This is very similar to the human

process for evaluating an image.

Ashken (2016) summarized the similarites between human vision and computer vision as

follows: “in both human and computer vision technology, the signal is understood through

comparison with a set of references. Human and computer vision systems then evaluate whether

[it is] sufficiently similar to be another example of the same thing.” Differences in the mode of

signal transmission (biological versus electrical signal) mean that computer vision is much faster

than human vision. Interestingly, interpretation first begins to take place in the human retina

whereas the computer sensors that capture light do not perform any analysis (Ashken, 2016).

Figure 3, below, is an example of a check image. The item is an altered money order and

can be immediately identified as fraudulent upon visual examination by a human analyst.

Figure 3. Example of an altered money order that is recognizable by the trained human eye
based on handwritten information (1) courtesy amount written and (2) legal amount written
which are both known to be typed on genuine items based on knowledge of (3) this type of
money order sold by King Soopers City Market for Western Union.

31
Just as a machine learning algorithm can facilitate the development of computer vision after

being shown a sufficient amount of training material, so can a human employee learn to

recognize counterfeit and altered instruments after learning their templates. Training is

recognized by the ACFE to be one of the most important components of detecting fraudulent

checks (Association of Certified Fraud Examiners, 2017, p. 407). Figure 4 is an example of a

valid postal money order with features annotated which could be used for training to detect

fraudulent items.

Figure 4. Example of United States Postal Money Order with 4 features indicated for use in
fraud detection: (1) serial number (2) security foil stripe (3) courtesy amount and (4) amount,
date, and location of sale. Reprinted from Domestic Money Orders by the United States Postal
Service, 2018, Retrieved from https://www.usps.com/shop/money-orders.htm. Copyright 2018
by USPS.

Theoretically, computers could be taught to recognize the same features and any

deviations through machine learning. Sohangir, Wang, Pomeranets, & Khoshgoftaar (2018)

noted that light, object shapes, and object materials can be further discerned by deep learning,

which is a reapplication of neural network machine learning (p. 2) Recent developments in

32
training techniques and computer vision suggest growth and broader applications of computer

vision are possible.

One company has already made an attempt. Cognizant Digital Business used a machine

learning model derived from Google TensorFlow to identify counterfeit checks by comparing

them to a consortium database of previously accepted ones. Similar to the block analysis

mentioned above, the solution used optical character recognition to “read” the check with a

digital eye and then parsed the information with a neural network. Cognizant Digital Business

helped their client bank analyze massive data sets collected from scanned checks—payee, check

number, account and routing numbers, amounts, endorsements, and maker signatures. From

these, the company established a set of normative features of good checks. Once the neural

network learned the rules, it continued to adapt based on new checks it was shown. The model

provided a confidence score on each check in 70 milliseconds or less, with responses ranging

from good to certain fraud. The bank determined responses to employ automatically regarding

whether funds would be released to the customer and how much, depending on the algorithm’s

rating. The solution boasted a $20 million projected reduction in fraudulent transactions and low

cost as a result of reduced manual effort (Cognizant, 2018)

There are some compliance limitations and concerns regarding the use of artificial

intelligence and machine learning. Financial institutions relying on artificially intelligent

decisions will struggle to prove they are not discriminatory against any protected class.

Fernandez (2017) discussed that some decisions create substantial operational, financial, and

reputational risk. While smart machines can automate tasks and machine learning models using

advanced statistical techniques can process far more transaction records than a human could,

there are still situations that require human judgment. These situations arise as financial
33
institutions navigate complex and nuanced compliance issues in a highly regulated industry.

Another issue is the feasibility of collecting and processing useful data from many disparate,

outdated systems and locating qualified engineers to manage the deployment of advanced

machine learning. For these reasons, one conclusion is that AI is not yet capable of supplanting

the human element in transaction reviews (Fernandez, 2017).

Summary

In summary, this study reviewed the history of checks, examined the current methods and

status of check fraud in 2018, and discussed current and possible future methods for preventing

check fraud. A form of payment with a history over a millennium long is unlikely to disappear

tomorrow. Quinn & Roberds (2008) noted the versatility and legal certitude of the check have

contributed to its persistence. In addition, checks are still reasonably cheap, accessible, and

widely accepted. There are some situations, such as real estate transactions and payments

between acquaintances, that still have few viable alternatives (p. 25). The Federal Reserve Bank,

which centrally clears all checks circulated in the United States, does not record return or loss

information. Therefore, the best estimates of the size and scope of check fraud emanate from

victims who self-report losses, and law enforcement, which reports activity detected and

prevented.

The most vulnerable victims generally possess the least ability to protect themselves from

scams. Individuals with limited numeric skills and more debt than they believed they could

handle were both more likely to be victims of fraud (Federal Trade Commission, 2013). Victims

usually send the funds by wire transfer, gift card, cash, or similar methods that are difficult to

cancel and/or trace. Some customers intentionally defraud their financial institution through

34
various forms of check fraud. This leaves the prevention of check fraud down to financial

institutions in the minutes surrounding the point of deposit; otherwise it is too late.

Financial institutions already utilize applications of AI, including machine learning, for

fraud transaction monitoring. As discussed by FINRA regarding their implementation of

machine learning to predict stock market fraud, machine learning could note factors significantly

indicative of fraud that humans would never have noticed. Still, humans are going to be a

necessary part of the process for the foreseeable future. Human analysts are needed to understand

the rationalizations of other humans and to control and review alerts generated by machines

(Financial Industry Regulatory Authority, 2018). Numerous studies of machine learning

applications were discussed in search of methods that may have relevance to the problem of

check fraud. Check fraud costs consumers and financial institutions a combined billion and some

dollars each year; however, information collected in this research regarding the weaknesses of

current methods and potential future directions will inform check fraud investigators.

Discussion of the Findings

This study reviewed scholarly documentation to gain a thorough understanding of checks,

check fraud, characteristics of fraud victims, and machine learning technologies. This research

study also examined the capabilities of both traditional check fraud detection techniques and

machine learning, discussing their combined uses in the detection of unusual activity. The study

lends legitimacy to the field of check fraud investigation by exposing the current size and scope

of the problem. As objectively as possible, relevant literature was located to address the problem

and several significant findings were discovered.

35
Major findings

The literature reviewed for this research ranged from survey reports from government

agencies and studies published in academic journals to marketing materials and references for

industry experts. The literature reviewed is dated between 1990 and 2018. The diverse literature

was selected for the breadth of sources and depth of focus to the subjects. It allowed for the

exploration of the different ways in which checks are used to defraud consumers and financial

institutions. The literature also showed the different measures currently in use by financial

institutions. In order to explore the different capabilities and uses of machine learning, it was

necessary to review literature from studies that concerned exploration at the cutting edge. This

variety of literature brought out the answers to the questions posed at the outset of the study.

Who is involved in check fraud and how does it take place? How could machine learning be used

to combat check fraud? Is it superior to—or could it supplement—existent solutions such as

behavioral alerts and consortiums?

First, a profile of check fraud was developed. The literature showed that victims of check

fraud can be any age, though people aged 20-29 were the most likely of any age group to be

victims of check fraud (Baker, 2018). The elderly and destitute are more seriously damaged by

fraud, even by the small dollar losses typical of check fraud. The majority of the check fraud

resulting in losses to consumers relies on the cooperation of victims, who provide their login

credentials and assistance returning the money. Perpetrators of check fraud use a variety of

schemes from employment to romance to lottery to induce victims to advance a fee from the

funds financial institutions make available following check deposits. Victims are left responsible

for repaying the funds sent to scammers when the check is returned, and consumers as well as

financial institutions collectively lose billions of dollars each year as a result of check scams.
36
 Secondly, the capabilities of artificial intelligence were explored. Applications of AI

including machine learning are already used to detect and predict fraud as well as other unusual

activity. Artificial intelligence is designed to mimic and improve upon the human brain’s ability

to process data. Thus, applications of AI can determine patterns and detect trends that are too

complicated for humans or other computational techniques to recognize. Computer vision and

text reading abilities are in use to “read” information both printed on the MICR line and

handwritten on checks. It is possible such technologies could soon be used to recognize

deviations (or other signs of alteration) on a check or money order.

In regard to the third question of whether the current measures are sufficient to mitigate

check fraud, the literature used in this research appears to illustrate that the current technology

measures used for preventing and detecting check fraud are rapidly becoming insufficient. With

check fraud losses appearing to surge in recent years, and the dollar volume of commercial

checks processed continuing to grow (Board of Governors of the Federal Reserve System,

2018b), analysts will soon be unable to keep up with the number of alerts. False positive rates

impact the effectiveness of a financial institution’s anti-fraud program. However, based on

machine learning breakthroughs, certain characteristics can be leveraged to help lower false

positive rates and increase the efficiency of detection. Machine learning does not provide a

shield against every fraud scheme, but it does provide a wide range of diverse applications and

techniques that can supplement and improve the accuracy of existing methods including fraud

software and consortiums.

Comparison of the findings with existing studies

The problem of check fraud is not new. In fact, it is so ancient that any study of check

fraud must be considered within the context of existing techniques to combat it. Manual review
37
of items is the simplest solution. This method is often successful in detecting alterations and

counterfeit items. Financial institutions also created manual review accounting strategies like

positive pay, which is designed to catch unauthorized items. Another manual strategy for

approaching the problem of check fraud is reliance on frontline staff training and experience.

Tellers recognize red flags in customer situations as well as spot and refuse suspicious items.

Gradually, more automated approaches have also been developed to accompany the continued

deployment of manual strategies.

Through contracts with software vendors or through developing in-house software

solutions, many financial institutions have implemented a system for fraud alerts. Some have

joined consortiums to query data from collective reports of fraudulent activity. While these

solutions are more effective than manual reviews alone, there are still significant numbers of

false positive alerts. Consortium data may be out of date, inaccurately reported by another

member, or lack detail due to sharing restrictions. Software alerts may trigger for benign activity

that is similar to fraud, or for transactions that exceed parameters set by the institution even

though they are legitimate. Unreliability of alerts and data are significant obstacles to catching

fraud. Any delay in review of alerts, if they are not real-time, also results in greater likelihood for

financial losses. Some will argue maintenance and refinement of these techniques is all that is

required to sufficiently mitigate check fraud.

This study was completed through academic research also within the context of available

literature. Some studies presented contrasting evidence with the above findings. One issue raised

by experts in regard to machine learning is that of scalability. As described in the study by

Wedge et al. (2017), machine learning has long been thought to require substantial computing

resources in order to evaluate the massive amounts of transactional data maintained by financial
38
institutions. Thus, a potential limiter for the use of artificial intelligence and machine learning is

resources. With sufficient training features and periodic historical data aggregation (rather than

constant), machine learning can be computationally economical. Finding and affording adequate

talent to employ continues to limit organizations from undertaking the implementation of

machine learning.

Furthermore, the financial benefit is relevant to the implementation of new fraud control

measures. Financial institutions already prevent the majority of fraud attempts according to the

American Banker’s Association. Law enforcement like the United States Postal Inspectors

intercept many more. Quantifying fraud that did not occur is difficult, which makes justification

of new technology onerous. Cognizant (2018) and Wedge et al. (2017) approached this obstacle

by testing their solutions on old data and projecting savings based on what would have been

detected.

One may also question the utility of the findings of this study based on several other

aspects of the topic. First, the value may be questioned due to the decline in check use both in the

United States and worldwide. Investment into a dying form of payment may not reap sufficient

returns. However, as noted above, the Federal Reserve data on commercial check clearing

suggests a recent plateau (Board of Governors of the Federal Reserve System, 2018b).

Concurrently, the FTC, FBI, National Consumers League, and BBB all received an increased

number of complaints regarding check fraud from 2014 to 2017 (Baker, 2018). Substantial

financial and psychological harm to fraud victims was also discussed. Existing studies

demonstrate more severe monetary losses result from types of fraud other than check fraud. As

well, it can be argued that victims of violent crime like rape and murder are more seriously

harmed than victims of white-collar crime.

39
 Finally, it is necessary to consider compliance issues when employing machine learning

techniques for any business process. This is particularly true in financial institutions, which are

highly regulated. Fernandez (2017) insinuated that regulators are not prepared for the use of AI

and machine learning—thus limiting the feasibility of such solutions for banks and credit unions

that fear regulator reprisal. Until sensitive data and decisions can be trusted with machines, and

legislation catches up to technology, humans will be a compulsory part of the process.

Limitations of the study

The first limitation of this research is due to the lack of availability of newer material.

While there are still relevant publications about checks from 2013 to 2018, some materials date

from 1990 to 1999. This research needs to acknowledge technological advances in the check

product, namely digital processing of checks through the federal reserve system and Remote

Deposit Capture, as well as compensate for the age of the volume of research on checks. Much

present-day research regarding financial transactions found is focused on virtual currency and

other electronic transactions.

Another limitation is a result of the reluctance of victims to acknowledge their own

victimization and losses. Fraudsters exploit psychological tricks, causing victims to comply with

demands and sometimes disbelieve that funds were taken. Some victims deny outright that they

were scammed due to embarrassment or even memory loss. With low reporting rates to the

government, law enforcement, and other agencies, a precise view of the issue cannot be gained.

Third, this study was limited by the lack of other research specific to the topic. While

both machine learning and checks have separately been thoroughly investigated, there is very

little information available regarding the combination of the two. Wedge et al. (2017) were

concerned with card fraud. Julisch (2010) referenced the broader topic of payment fraud. Moula
40
et al. (2017) focused on credit default prediction. Lidestri (2018) primarily researched

cybersecurity. Each of these studies concern adjacent subjects to check fraud but are not exactly

the same topic. An extensive search was conducted and only Cognizant (2018) was found to

have deployed machine learning to recognize counterfeit checks by “looking at” check images

during the time of deposit. Competing techniques were not located and this may weaken the

study’s ability to evaluate this approach.

This study was also limited by obtainability and nationality of references. The study is

comprised of books, papers, and articles available on the open Internet as well as the library

databases accessible through Utica College. Some databases and potentially relevant studies

could not be accessed due to a lack of subscription. The study was U.S.-focused and was

therefore biased toward targeting predominantly American references.

The fifth—and perhaps most significant—limitation to this research has been the lack of

access to confidential or proprietary material. Financial institutions, fearing damage to their

reputations and profitability, do not widely publicize their losses. Nor, due to financial privacy

laws, can they disclose any consumer or suspect information. Further, for-profit businesses

wishing to maintain their edge over competitors seldom share their insider secrets, including

anti-fraud practices. Law enforcement agencies cannot release information pertaining to open

investigations. Lawyers, who may possess information from fraud offenders, are bound not to

release information by attorney-client privilege. These restrictions limited the scope of the study

to open source materials pertaining to the topics explored. The limitations of this research may

impact recommendations and conclusions, which must inherently be based on incomplete

information.

41
 Recommendations

Based on the findings of this research, additional research and action to combat check

fraud is suggested. It is clear that the check as a form of payment is not yet going extinct.

Consumers require financial institutions to act to protect them from victimization, and better

protection from check fraud will also benefit the financial institutions in terms of strengthened

profits, reduced losses, and improved customer relationships. Given advancements in artificial

intelligence including machine learning, these ends certainly appear possible. Recommendations

for action are therefore enumerated below.

Recommendations for future research

This study showed that technological capabilities have advanced in recent years. Julisch

(2010) acknowledged the pointlessness of fraud detection programs that merely alert to fraud

that has already happened rather than anticipate or prevent new fraud. Innovation is required to

better protect individuals and financial institutions from check fraud. Computer vision and other

AI techniques, such as machine learning, could allow near instantaneous visual review of checks

as well as the ability to make availability decisions with a human-like eye. More research is

needed into machine learning and computer vision solutions as they specifically apply to check

fraud.

This research asked: “How could machine learning be used to combat check fraud? Is it

superior to—or could it supplement—existent solutions such as behavioral alerts and

consortiums?” Further research is required to determine in what capacity machine learning can

be leveraged for fraud investigation, prevention, and detection. Solutions like the one put forth

by Cognizant (2018) must be tested in an objective manner. Experimentation should also be

performed with the algorithms most applicable to the problem of check fraud identified above:
42
random forests (Wedge, Kanter, Veeramachaneni, Rubio, & Perez, 2017), support vector

machines (Moula, Guotai, & Abedin, 2017), and neural networks (Krizhevsky, Sutskever, &

Hinton, 2012; Cognizant, 2018).

Some best practices were identified in this research. Ideally, algorithms should be trained

with a sufficient sample size of training data as in Krizhevsky, Sutskever, & Hinton (2012) and

Moula, Guotai, & Abedin (2017). Not only should there be an extensive amount of training data,

it also should be selected with care to allow the most accurate behavioral analysis (Wedge et al.,

2017). Data should be pre-processed to standardize image size and other factors to increase

accuracy (Wu, Hu, & Huang, 2014; Krizhevsky, Sutskever, & Hinton, 2012). Supervised and

reinforcement methods of training machine learning models are recommended to allow control

for desired results and adaptation over time (Jean-Philippe, 2018). These measures should be

considered during any further experimentation.

Several moral dilemmas were raised by this research. The first is the intersection of

compliance and artificial intelligence. How will financial institutions prove to regulators and

customers that their machine learning algorithms are not discriminating against certain races,

genders, other minorities, or protected classes? Fairness in lending, check holds, and all other

financial institution policies must be maintained at all times. Secondly, selling anti-fraud

technologies also presents ethical quandaries. Software companies act as hired mercenaries

locating fraud for financial institution clients, who pay them hundreds of thousands of dollars.

Banks and credit unions seem to have only two other alternatives—developing software in-

house, or bleeding out money to fraudsters instead. Consumers have very little knowledge about

or control over their financial institution’s fraud prevention program and may literally pay the

difference for any gaps in protection. Even if it were possible, software companies do not truly
43
wish to irradicate all fraud or they would lose their source of income. How much responsibility

do financial institutions and software companies bear for protecting victims from fraud?

Additional literature research should also be undertaken to build upon what was

discovered in this research and improve upon the limitations mentioned above. Check fraud

should be reviewed globally to determine the scale of check fraud worldwide. Trends and

recommendations from other countries could be illustrative to investigators combatting check

fraud in the United States. Additionally, relevant machine learning studies concerning fraud and

check fraud may have been completed internationally, published in journals not accessible

through Utica College or the open Internet, or otherwise overlooked by this research. Last, the

lack of recently dated material created a gap in this research. Further work should be done to

locate contemporary studies of checks, including newer digital services such as remote deposit

capture and electronic conversion. Broadening the scope of research on both machine learning

and check fraud will further serve to educate investigators, legislators, and consumers.

Recommendations for combatting check fraud

In addition to research, practical recommendations for banks, credit unions, and other

financial organizations were identified. Financial institutions and consumers alike are harmed by

the lack of real-time, widespread, and effective check fraud prevention tools. Far from dying off

as a fossil form of payment from the 15th Century, checks have maintained their relevance for

both individuals and businesses for situations without a suitable substitute form of payment.

While some cutting-edge techniques remain at least a few years from deployment at most

companies, there are some actions that can be taken now to significantly decrease losses from

check fraud.

44
 Financial institutions should participate in consortiums and other information-sharing

programs as permissible by law. While fraudsters freely share tips and tricks for committing their

crimes online, financial institutions are typically characterized by their secrecy. Though financial

institutions are bound by regulations regarding private information of consumers, consortium

information strengthens anti-fraud programs and can be accessed within compliance confines.

Cross-institution information significantly undercuts fraudsters’ ability to repeat the same scam,

deterring crime (Castanheiro, 2017). Consortiums can be queried in real time and responses from

the database can be incoporated into bank policies. Some examples include account opening and

deposit hold decisions. Weaknesses of consortiums were identified, including the potential for

outdated information and false positives. Therefore, consortiums must be deployed alongside

several other measures.

It is necessary to implement cross-channel, real-time fraud alerts that integrate behavior,

knowledge, and machine learning. Delayed alerts allow time for customers to withdraw funds

made available from check deposits. Further, analysts working alerts concerning activity that is

days or even weeks old are looking in the rearview mirror. Alerts that are too narrowly focused

on one type of payment can miss the full context of real fraudulent activity. Conversely, software

that risk-scores transactions and intelligently makes decisions regarding funds availability will

prevent losses and even allow identification of emerging trends. Financial institutions must

continually tune their software to avoid creating an avalanche of false positive alerts and, again,

a delayed review of activity.

Factors to improve predictive capabilities of the software should be both qualitative and

quantitative to allow for a spectrum of risk-taking as detailed in Matsatsinis (2002). Customer

financial status modeling appears to have direct applications to risk-rating immediate provision
45
of funds from check deposits. It is similar to providing the customer a short-term loan while the

check is collected upon. Quantitative data is abundant—credit score, account balances, years as a

customer—and provides useful information in scoring the inherent risk of the customer. Patterns

of fraud identified by the ACFE like frequent transactions and unusual deposit behavior, will

need to be built into the system (Association of Certified Fraud Examiners, 2017, pp. 434-435).

Financial institutions may tailor certain dollar thresholds that are deemed to be low risk as well

as set “normal” behavior, allowing for controlled alerts within the financial institution’s unique

approach and customer base. Financial institutions also hold significant amounts of qualitative

data that can be used to identify risk factors for fraud victimization. Occupation and marital

status may indicate an extra vulnerability to fraud (Pak & Shadel, 2011). All customer profile

data points should be included in analysis along with the above. Robust software that generates

real-time alerts with these considerations should be pursued.

Consumers should also take responsibility for their own protection against forms of

fraud, including check fraud. Members of the public should take caution to verify the legitimacy

of individuals and companies they transact with, especially online. If in doubt about a situation,

consumers should slow down and reach out to their financial institution’s fraud department or

research further before acting. Online banking credentials should never be shared with a third

party such as a friend, employer, or lawyer. Checks should not be accepted from unsolicited

mailings, or for more than the selling price of an item listed for sale. When a situation

necessitates the receipt of a check as payment, consumers should be vigilant for any red flags in

the sender’s communication or on the item itself: misspellings, distorted logos, or blurry

signatures, for example. Consumers should refrain from withdrawing funds immediately from a

check deposit, even if their bank makes funds available, and should never agree to send a portion
46
of the funds back to the sender of the check. Customers of financial institutions should regularly

review their account statements for any unauthorized activity and report it promptly as directed

in their account agreements. If consumers take charge of their own protection from scams, they

will present undesirable targets to scammers and become much safer.

It is also time for a collective education effort as was made in 2009. Financial institutions

and law enforcement must again drive public awareness of check fraud. Media campaigns should

be launched detailing the common patterns and relevant laws. Websites that facilitate online

dating, classifieds, and money transfers should also undertake this effort. Government and non-

profit programs for the elderly and underprivileged must strive to include education regarding

scams that target their clients. Financial institutions must create comprehensive programs to

educate their staff on signs of fraudulent items, red flags in customer behavior, and best practices

for educating customers. Many banks could better explain aspects of banking that lead to

misconceptions.

Ignorance allows the losses resulting from check fraud. One example requiring more care

is the requirement for funds released under the Regulation CC schedule. The check clearing

process is also not often understood by customers. Additionally, financial institutions and

regulators alike should attempt to bring more publicity to the bounds of coverage by federal

deposit insurance. Last, it is important for consumers to be informed regarding liability for

returned items. Education is the only solution that empowers consumers to make better decisions

about accepting checks from strangers, friends, and online acquaintances.

The last recommendation for combatting check fraud is to perform workforce planning

and vendor evaluation. According to expert projections as in Lidestri (2018), the volume of

activity at organizations throughout the country will soon outpace the ability of current staff to
47
review. It is imperative that financial institutions employ automation wherever possible to

preserve human efforts for situations where they are truly needed. Commiserate with their size,

banks and credit unions will also need to employ a staff of analysts, investigators, and experts

with adequate credentials to investigate check fraud and other cases. Finally, and perhaps most

importantly, financial institutions will need qualified software engineers to create, test, and

supervise machine learning solutions for fighting fraud. Most firms will need to begin efforts

now to justify new positions or vendor contracts to management.

Conclusion

For over six centuries, checks have ruled the payment landscape: overcoming distance,

surviving wars, and shaping the establishment of the modern centralized Federal Reserve

banking system. Checks are a secure and convenient instrument for ordering the transfer of funds

between parties. Day-to-day purchases have been assumed by methods like the debit card and

credit card, which recently became much more secure when with the transition to EMV

technology in the United States. Innovation also continues in the realm of new electronic transfer

services and currencies but, short of RDC and electronic conversion services, the check has

remained relatively unchanged for the last 50 years.

While a profile of those with a propensity to be victimized by check fraud has been

developed through this research, individuals of any age, race, or income can fall prey to these

schemes. The more concerning truth is that not everyone is equally harmed by the resulting loss.

The most financially and mentally vulnerable have far more to lose in a single risky transaction,

even if a smaller portion of funds are ultimately stolen. Perpetrators of check fraud—such as

counterfeiters working out of West Africa—are difficult to locate and even more difficult to

bring to justice. Financial institutions must either pursue the customers that warranted the check
48
in the deposit process or sustain the loss. Therefore, the financial institution’s duty for preventing

fraud is both fiduciary and self-preserving in nature.

The key to recognizing and blocking the majority of check fraud may be hidden within

massive amounts of transaction data, unrecognizable by human analysts who are only able to

view a tiny portion of the picture. Methods like data mining and cross-institution collaboration

help reveal information. However, they must be used more widely and enhanced by modern

techniques to keep up with the overall volume of banking activity. Artificial intelligence,

including applications like machine learning, data mining, and computer vision can deliver

objective results, inform rule-making, and recognize trends without overburdening analysts.

Machine learning should be seen as an enhancement—not a substitute—for human

wisdom and experience. Machines still lack the ability to understand human emotions,

rationalizations, and choices. Despite huge advancements in software algorithms, machines

cannot be entirely autonomous in decision making in the financial industry. Oversight is required

when updating models, assuring compliance with applicable rules and regulations, and

evaluating unique cases. Human intelligence will still be indispensable in the process, although

increasingly in a more sophisticated role as automation is developed.

Criminals who commit check fraud rely on the isolation of information for their crimes to

persist. They bank on the ignorance of new potential victims. The seclusion of financial

institutions who refrain from information sharing and silence of victims who avoid reporting

allow scammers to repeat the same schemes again and again. Until the check is completely

relegated to museums, law enforcement, financial institutions, legislators, and consumers alike

must work together to bring information about check fraud forward. Through collaboration,

education, and leveraging new tools, consumers can be better protected from check fraud.
49
 References

Agarwal, A., Gupta, A., & Hussein, K. (1997). Bank check analysis and recognition by

computers. In H. Bunke, & P. S. Wang, Handbook of character recognition and

document image analysis (pp. 623-651). River Edge: World Scientific Publishing

Company.

Akinladejo, O. H. (2007). Advance fee fraud: trends and issues in the Caribbean. Journal of

Financial Crime, 320-339.

American Bankers Association. (2018, January 24). Banks stop $17 Billion in fraud attempts in

2016. Retrieved from https://www.aba.com/Press/Pages/012418DDAFraud.aspx

Archer, A. K. (2017). "I made a choice": Exploring the persuasion tactics used by online

romance scammers in light of Cialdini's compliance principles. Retrieved from

https://epublications.regis.edu/cgi/viewcontent.cgi?article=1830&context=theses

Ashken, S. (2016, August 15). What is computer vision? Part 4: Human vs. computer vision.

Retrieved from https://www.blippar.com/blog/2016/08/15/what-computer-vision-part-4-

human-vision-vs-computer-vision

Association for Financial Professionals. (2018, April). 2018 payments fraud and control survey

report: Key highlights. Retrieved from

https://commercial.jpmorganchase.com/jpmpdf/1320745402134.pdf

Association of Certified Fraud Examiners. (2017). Fraud examiners manual. Austin: Association

of Certified Fraud Examiners.

50
Baker, S. (2018, September 15). Don't cash that check: BBB study shows how fake check scams

bait consumers. Retrieved from https://www.bbb.org/en/us/article/news-releases/18367-

dont-cash-that-check-bbb-study-shows-how-fake-check-scams-bait-consumers

Bank of America. (2018, June 8). Deposit agreement and disclosures. Retrieved from

https://www.bankofamerica.com/deposits/resources/deposit-agreements.go

Better Business Bureau. (2018). 2017 BBB scam tracker annual risk report: New trends in scam

risk. Retrieved from https://www.bbb.org/globalassets/local-bbbs/council-

113/media/scam-tracker/risk-report/bbbscamtrackerannualreport-final-2017.pdf

Board of Governors of the Federal Reserve System. (2016, June 16). Check Services. Retrieved

from https://www.federalreserve.gov/paymentsystems/check_about.htm

Board of Governors of the Federal Reserve System. (2017, March 1). Compliance guide to small

entities Regulation CC: Availability of funds and collection of checks. Retrieved from

https://www.federalreserve.gov/supervisionreg/regcccg.htm

Board of Governors of the Federal Reserve System. (2018a, July 23). 2016 Survey of consumer

finances. Retrieved from https://www.federalreserve.gov/econres/scfindex.htm

Board of Governors of the Federal Reserve System. (2018b, February 15). Commercial checks

collected through the Federal Reserve--Annual data. Retrieved from

https://www.federalreserve.gov/paymentsystems/check_commcheckcolannual.htm

Castanheiro, M. (2017, October 4). Fraud trends targeting credit unions & members [Webinar].

Retrieved from https://verafin.com/resource/fraud-trends-targeting-credit-unions-

members/

51
Chatzis, S. P., Siakoulis, V., Petropoulos, A., Stavroulakis, E., & Vlachogiannakis, N. (2018).

Forecasting stock market crisis events using deep and statistical machine learning

techniques. Expert Systems with Applications, 112, 353-371.

Check Fraud Working Group. (1999, February). Check fraud: a guide to avoiding losses.

Retrieved from https://www.occ.treas.gov/publications/publications-by-type/other-

publications-reports/pub-other-check-fraud.pdf

Cliff, G., & Wall-Parker, A. (2017, April). Statistical analysis of white-collar crime. Retrieved

from Criminology and Criminal Justice:

http://criminology.oxfordre.com/view/10.1093/acrefore/9780190264079.001.0001/acrefo

re-9780190264079-e-267

Cognizant. (2018). Advanced AI machine learning solution detects check fraud for a large global

bank. Retrieved from https://www.cognizant.com/case-studies/pdfs/ai-driven-solution-

reduces-fraud-risk-for-bank-codex3688.pdf

Department of the Treasury. (2009, June 16). Guidance on the Scope of Permissible Information

Sharing Covered by Section 314(b) Safe Harbor of the USA PATRIOT Act . Retrieved

from https://www.fincen.gov/sites/default/files/shared/fin-2009-g002.pdf

Dua, S., & Du, X. (2011). Data mining and machine learning in cybersecurity. Boca Raton:

Auerbach Publications.

Early Warning. (2015, May 12). Early warning and NCR to enable cross-institution real-time

good funds network. Retrieved from https://www.earlywarning.com/news/press-

releases/2015/early-warning-and-ncr-to-enable-cross-institution-real-time-good-funds-

network.html

52
Executive Office for United States Attorneys. (2017). United States attorneys' annual statistical

report fiscal year 2017. Retrieved from

https://www.justice.gov/usao/page/file/1081801/download

Federal Bureau of Investigation. (2017). 2017 Internet Crime Report. Retrieved from

https://pdf.ic3.gov/2017_IC3Report.pdf

Federal Bureau of Investigation. (n.d.). Internet crime schemes. Retrieved from

https://www.ic3.gov/crimeschemes.aspx

Federal Deposit Insurance Corporation. (2018, January 1). A guide to what is and is not

protected by FDIC insurance. Retrieved from

https://www.fdic.gov/deposit/covered/notinsured.html

Federal Trade Commission. (2013, April). Consumer fraud in the United States, 2011 the third

FTC survey. Retrieved from

https://www.ftc.gov/sites/default/files/documents/reports/consumer-fraud-united-states-

2011-third-ftc-survey/130419fraudsurvey_0.pdf

Federal Trade Commission. (2018, March). Consumer Sentinel Network Data Book 2017.

Retrieved from https://www.ftc.gov/system/files/documents/reports/consumer-sentinel-

network-data-book-2017/consumer_sentinel_data_book_2017.pdf

Fernandez, D. (2017, September 20). Is artificial intelligence ready for financial compliance?

Retrieved from https://www.bloomberg.com/professional/blog/artificial-intelligence-

ready-financial-compliance/

Financial Crimes Enforcement Network. (2018). Suspicious activity report statistics (SAR stats).

Retrieved from https://www.fincen.gov/reports/sar-stats

53
Financial Industry Regulatory Authority. (2018, July 17). How the cloud and machine learning

have transformed FINRA market surveillance [Podcast]. Retrieved from

http://www.finra.org/industry/podcasts/how-cloud-and-machine-learning-have-

transformed-market-surveillance

FLEX Credit Union Technology. (n.d.). FLEX TrueChecks Integration. Retrieved from

https://info.flexcutech.com/hubfs/FLEX%20TrueChecks%20Integration-

1.pdf?submissionGuid=7e4a0361-ef39-4d61-9b9a-ebd4a7254ce1

Ganzini, L., McFarland, B., & Bloom, J. (1990). Victims of fraud: Comparing victims of white

collar and violent crime . The Bulletin of the American Academy of Psychiatry and the

Law, 55-63.

Guardian Analytics. (2016, March). Mobile banking fraud trends: Fraudulent checks meet

remote deposit capture. Retrieved from https://guardiananalytics.com/wp-

content/uploads/2018/06/FraudUpdate_MobileRDCTrends_March2016.pdf

Hoffmann, A. O., & Birnbrich, C. (2012). The impact of fraud prevention on bank-customer

relationships. The International Journal of Bank Marketing, 390-407.

Jean-Philippe, R. (2018, May). Enhancing computer network defence technologies with machine

learning and artificial intelligence (capstone project). Utica, New York: Utica College.

Julisch, K. (2010, September 9). Risk-based payment fraud detection. IBM Research, pp. 1-13.

Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet classification with deep

convolutional neural networks. Advances in Neural Information Processing Systems 25.

NIPS.

Kutler, J. (2017, March). Financial regulators embrace artificial intelligence. Institutional

Investor.
54
Legal Information Institute. (n.d.). Bank fraud. Retrieved from

https://www.law.cornell.edu/uscode/text/18/1344

Lidestri, N. (2018, May). The impact of artificial intelligence in cybersecurity (capstone project).

Utica, New York: Utica College.

Manhattan District Attorney's Office. (2018, July 11). DA Vance: ID thieves indicted for “Card

Cracking” scheme involving Snapchat, rent checks. Retrieved from

https://www.manhattanda.org/da-vance-id-thieves-indicted-for-card-cracking-scheme-

involving-snapchat-rent-checks/

Matsatsinis, N. F. (2002, May). An intelligent decision support system for credit card assessment

based on a machine learning technique. Operational Research, 2(2), 243-260.

Mazur, M. (2014, January). Fraud grows, but new tools help check it. ABA Banking Journal, p.

11.

Monga, V. (2014, March 10). U.S. companies cling to writing paper checks. Retrieved from The

Wall Street Journal: https://www.wsj.com/articles/u-s-companies-cling-to-writing-paper-

checks-1394494772

Moula, F. E., Guotai, C., & Abedin, M. Z. (2017). Credit default prediction modeling: an

application of support vector machine. Risk Management, 158-187.

National Consumers League. (2018, February). Top scams of 2017. Retrieved from

https://d3n8a8pro7vhmx.cloudfront.net/ncl/pages/4567/attachments/original/1517428302

/2017_top_scams_report.pdf?1517428302

Pak, K., & Shadel, D. (2011, March). AARP Foundation national fraud victim study. Retrieved

from https://assets.aarp.org/rgcenter/econ/fraud-victims-11.pdf

55
Qiao, Q., & Beling, P. A. (2016, June). Decision analytics and machine learning in economic and

financial systems. Environment Systems & Decisions, 36(2), 109-113.

Quinn, S., & Roberds, W. (2008). The evolution of the check as a means of payment: A historical

survey. Retrieved from https://www.frbatlanta.org/-

/media/Documents/research/publications/economic-

review/2008/vol93no4_quinn_roberds.pdf

Rouse, M. (2014, August). False positive. Retrieved from

https://whatis.techtarget.com/definition/false-positive

Salinger, L. M. (2005). Encylopedia of white-collar & corporate crime, volume 1. Thousand

Oaks: Sage.

Silva, V. G., Ramalho, E. A., & Vieira, C. R. (2017). The use of cheques in the European Union:

A cross-country analysis. Open Economies Review, 581-602.

Smith, C. W. (2001, June). Defense to a payor bank's liability for late returns. CCH Deposit Law

Notes.

Sohangir, S., Wang, D., Pomeranets, A., & Khoshgoftaar, T. M. (2018). Big data: Deep learning

for financial sentiment analysis. Journal of Big Data, 1-25.

Telang, R., & Somanchi, S. (2016, October). Security, fraudulent transactions and customer

loyalty: A field study. Retrieved from

https://www.ftc.gov/system/files/documents/public_comments/2016/10/00062-

129181.pdf

United States Postal Service. (2018). Domestic Money Orders. Retrieved from

https://www.usps.com/shop/money-orders.htm

56
Verafin. (2017a, March). Fraud schemes & scams. Retrieved from https://verafin.com/wp-

content/uploads/2017/03/fraud-schemes-scams-IG-Verafin-170329.pdf

Verafin. (2017b, December 19). Unmasking online account takeover. Retrieved from

https://verafin.com/2017/12/unmasking-online-account-takeover/

Wedge, R., Kanter, J. M., Veeramachaneni, K., Rubio, S. M., & Perez, S. I. (2017, October 20).

Solving the “false positives” problem in fraud prediction: Automated data science at an

industrial scale. Retrieved from http://arxiv-export-lb.library.cornell.edu/pdf/1710.07709

Westphal, C. (2009). Data mining for intelligence, fraud, & criminal detection. Boca Raton, FL:

CRC Press.

Wu, H.-C., Hu, Y.-H., & Huang, Y.-H. (2014). Two-stage credit rating prediction using machine

learning techniques. Kybernetes, 43(7), 1098-1113.

57