Food Control 131 (2022) 108440

Contents lists available at ScienceDirect

Food Control
journal homepage: www.elsevier.com/locate/foodcont

A holistic approach to food fraud vulnerability assessment

Faeze Rezazade a, *, Jane Summers a, Derek Ong Lai Teik b
a
University of Southern Queensland, West St, Darling Heights, QLD, 4350, Australia
b
Sunway University, 5, Jalan Universiti, Bandar Sunway, 47500, Malaysia

A R T I C L E I N F O A B S T R A C T

Keywords: The increase in fraudulent incidents has resulted in a change of focus in theory related to investigating food fraud
Food fraud vulnerability from risk mitigation to vulnerability reduction. While the literature provided a framework from which to identify
Bayesian network possible areas of vulnerability to food fraud, it does not provide a methodology that would allow food producers
Decernis food fraud database
and processors to quantify and assess their vulnerability to food fraud more accurately. This paper developed a
Holistic approach
holistic approach to analyse food fraud vulnerability factors using a Bayesian Network (BN) approach based on
the Decernis food fraud database incidents. Food fraud incidents related to seafood, dairy, alcoholic beverages
and meat products were reviewed as they cover more than 50 % of the Decernis incident records (as in June
2018). 580 cases of food fraud were included in the development of the BN model. SPSS Modeler 18.2 was used
to construct two BN model, and fraud vulnerability factors were directly linked to seven criteria of the country of
origin, country of detection, year, food fraud types, product types, the weight of evidence, and types of adul­
terants. Possible food fraud vulnerability factors related to each case retrieved from Decernis database, combined
Barrier Analysis technique and Routine Activity Theory, and additional sources from the literature. Two Bayesian
Networks of Tree Augmented Naïve (TAN) and Markov were selected to determine the most reliable holistic
model. Based on the analysis result, the TAN model assessed the vulnerability to food fraud with a higher ac­
curacy rate of 86 %. The country of origin (76 %), food product types (9 %), types of adulterants (counterfeiting)
(9 %), and country of detection (China) (6 %) were the main predictors of food fraud vulnerability factors. This
model helps authorities in border protection, policymakers, and quality assurance agencies assess fraud
vulnerability for a range of food products for incoming (imported) food products, mainly if they know the
country of origin and type of food products.

1. Introduction
The failure of traditional food safety management systems in
detecting incidents of food fraud resulted in researchers calling for an
innovative approach to shifting the focus from risk mitigation to
vulnerability reduction (Everstine, Spink & Kennedy, 2013; Silvis et al.,
2017; Spink et al., 2017; Tähkäpää et al., 2015; Van Ruth, Huisman &
Luning, 2017). Vulnerability reduction seeks to eliminate the root causes
of food fraud that create opportunities and motivation for fraudsters
(Spink et al., 2016; Spink & Moyer, 2013). The shift in thinking iden­
tifies and assesses food fraud vulnerability factors (Van Ruth, Huisman
& Luning 2017) rather than implementing reactive countermeasures (e.
g. DNA analysis) once the fraud has occurred (Spink & Moyer 2011). The
innovative approach to undertake food fraud vulnerability assessment
should be based on an initial screening of historical incidents of food
fraud (Spink et al., 2016) as well as consideration of ways to identify and
assess the root causes of food fraud (Lord et al., 2017; Silvis et al., 2017;
Spink & Moyer, 2011; Van Ruth, Huisman & Luning, 2017).
Interest in food fraud vulnerability assessment is only just emerging,
despite the increased managerial interest in the issue of food fraud
prevention. (Silvis et al., 2017; Spink et al., 2019a, 2019b; Van Ruth
et al., 2018; Van Ruth, Huisman & Luning, 2017). Scholars currently use
one of two methods to undertake the food fraud vulnerability assess­
ment, the SSAFE tool and Carver + Shock method (Silvis et al., 2017;
Van Ruth et al., 2018; Wiśniewska, 2017; Yan et al., 2020).
The SSAFE tool comprises 50 questions based on the Routine Activity
Theory classification of opportunity, motivation, and countermeasures
factors as the root causes of food fraud (or vulnerability factors) (Silvis
et al., 2017; Spink & Moyer, 2013; Van Ruth et al., 2018; Van Ruth,
Huisman & Luning, 2017). It is a self-assessment tool for food businesses

* Corresponding author.
E-mail addresses: Faezeh.rezazade@gmail.com, faye.rezazade@excelsia.edu.au (F. Rezazade), Jane.Summers@usq.edu.au (J. Summers), dereko@sunway.edu.my
(D.O. Lai Teik).

https://doi.org/10.1016/j.foodcont.2021.108440
Received 28 March 2021; Received in revised form 14 July 2021; Accepted 19 July 2021
Available online 20 July 2021
0956-7135/© 2021 Elsevier Ltd. All rights reserved.
F. Rezazade et al. Food Control 131 (2022) 108440

Table 1
Main food fraud database features and limitations.
Database Food fraud classifications
RASFF (1) Improper, fraudulent, missing or
absent health certificates; (2) illegal
importation; (3) tampering; (4)
improper, expired, fraudulent or
missing common entry documents
or import declarations; (5) false
expiration date; and (6) mislabelling
EMA (1) Intentional distribution of
contaminated products; (2) artificial
enhancement; (3) counterfeiting;
(4) substitution; (5) mislabelling;
(6) dilution; (7) transhipment; and
(8) theft and resale.
Decernis (1) Replacement, (2) addition, (3)
removal
Limitations
➢ Geographic limitation: only
includes incidents in the EU
➢ Includes both intentional
and unintentional food
safety incidents
➢ Geographic limitation- Only
incidents in the USA
➢ Unavailable at the time of
data collection for this
research
➢ Not publicly available (only
through a paid subscription)
Decernis food fraud database was used as it references real-data cases of
global food fraud incident (Decernis,)s. Other databases such as RASFF
and EMA (see Table 1) provide information about food fraud incidents
that are geographically limited to European Union (RASFF) or the
United States (EMA). When combining these two databases, duplication
of incidents may exist as EMA also ‘collect [s] food fraud reports from
the RASFF’ database (Bouzembrak et al., 2018).
The data published in the Decernis database has also been reviewed
by an expert panel and academics who in turn have made recommen­
dations to ‘identify trends and vulnerabilities’ (Decernis). Each Decernis
incident record contains the following information about criteria of the
country of origin, country of detection, the weight of evidence (low,
medium, high), year, food fraud types, and adulterant types. The in­
formation related to each incident were collected and restored in an
Excel file.

Source: Bouzembrak and Marvin (2016); Manning and Soon (2016); Moore,
Spink and Lipp (2012), Zhang and Xue (2016).
to score the importance of these factors in their supply chain. The
Carver + Shock method is sometimes referred to as Vulnerability
Assessment Critical Control Points (VACCP) auditing scheme, and is
required by food safety regulators such as FDA (2009) and Global Food
Safety Initiatives (GFSI, 2018; Manning & Soon, 2016). Since food fraud
‘does not have a traditional terrorist shock value’ (Spink & Moyer, 2011,
p. R160) this method does not apply to food fraud vulnerability
assessment. Both of these afore mentioned methods rely on a manager’s
perspective, rather than on the initial screening of real-data required for
a comprehensive food fraud vulnerability assessment (Spink et al.,
2019b; Spink et al., 2016).
A holistic approach is required to consider a global food fraud
database, analyse food fraud vulnerability factors and their de­
pendencies. Studies by Marvin et al. (2016), Bouzembrak and Marvin
(2016), and Soon (2020) have shown the usefulness of the Bayesian
Network approach in linking factors influencing food fraud, their
interaction, with data from a global database (RASFF) to predict food
fraud incident types.
A Bayesian Network (BN) model approach developed in previous
research has shown successful prediction of more than 85% of food fraud
incident types when variables of the country of origin, product category,
country of detection, and year of the fraud occurring were known
(Marvin et al., 2016; Soon, 2020). Whilst, the aim of these studies was to
predict food fraud types, rather than to assess food fraud vulnerability,
the BN approach for this study is still relevant. Adopting a BN approach
will allow for modelling the dynamics and reciprocal interrelationships
between food fraud incidents and drivers of food fraud, or food fraud
vulnerability factors.
This paper takes up the call from the literature to develop an inno­
vative, holistic approach based on a BN model that links variables and a
global food fraud data source (Decernis,) to analyse the factors influ­
encing food fraud vulnerability.
2. Materials and methods
2.2. Identification of food fraud vulnerability factors
The food fraud vulnerability factor for each incident was determined
by literature (Silvis et al., 2017; Van Ruth et al., 2018), a global food
fraud database (Spink et al., 2016), and expert judgement (Marvin et al.,
2016). A Barrier Analysis Technique, combined with Routine Activity
Theory, was used to identify food fraud vulnerability factors in the
Decernis food fraud database. This research followed the process of
previous scholars (Marvin et al., 2016; Rezazade, 2020; Rezazade,
Summers & Ong, 2020) to identify vulnerability factors related to each
incident and index them based on the Routine Activity Theory classifi­
cations of opportunity, motivation, and countermeasure. Two external
reviewers (with expert knowledge) have reviewed and confirmed the
inclusion of food fraud vulnerability factors.
The data from the Decernis food fraud database and the results of the
Barrier Analysis Technique were stored in the Excel file to build the BN
model. The Excel file includes one column for each node. Each row in the
Excel file shows each incident case of food with all relevant data for the
various drivers (i.e., country of origin, food fraud types, food fraud
vulnerability factors).
2.3. Bayesian network modelling
A Bayesian Network is a graphical model and consists of the
following: (1) a set of random variables (or node) and a set of directed
links (or edges) between variables; (2) a set of mutually exclusive states
within each random variable which explain ‘the condition of a variable’
(Marvin et al., 2016, p. 465); and (3) ‘a directed acyclic graph’ (DAG)
formed from the variables and the directed links (or edges) (Jensen &
Nielsen, 2007; Marvin et al., 2016, p. 465; Tien and Der Kiureghian,
2016). A DAG is a directed graph with no directed path from A_i to A_j (e.
g. A_i→⋯→Aj) (Jensen & Nielsen, 2007). If there is a link between
Ai ​ to ​ Aj , then the random variable Ai is the parent of Aj and Aj is the
child of Ai (Jensen & Nielsen 2007; Marvin et al., 2016). A Bayesian
Network considers the joint probability distribution of all variables,
P(U) = P(A1 ,….., An ) ‘given by the product of all Conditional Probability
Tables’ (Jensen & Nielsen 2007, p. 36) as shown in Eq. (1):

The holistic approach developed for this study included the ∏

following four steps: (1) food fraud incident collection; (2) identification
of food fraud vulnerability factors through a combination of Routine
Activity Theory and Barrier Analysis technique; (3) constructing the
Bayesian Network model; and (4) validating the model.
2.1. Food fraud incident collection
Food fraud vulnerability assessment involved a review of each inci­
dent and its history (Spink et al., 2019b; Spink et al., 2016). The
n
P(U) = P(Ai| parents (Ai) ) (1)
i=1
Bayesian Networks also can calculate new probabilities when new
information about the state of a set of random variables is available
(Hossain & Muromachi, 2013; Jensen & Nielsen, 2007). The new in­
formation provides the evidence, meaning that ‘some of the variables are
observed’ taking values from their domains (Marvin et al., 2016, p. 466).
For example, if the evidence (e) related to the state of m variables
(e1, ​ ……, em ) becomes available, then inserting this new information to
Eq. (2) we can get Eq. (3):

2
F. Rezazade et al. Food Control 131 (2022) 108440

∏
n ∏m Table 2
P(U, e) = P(Ai |parents (Ai )) j=1 ej (2) Food fraud types in Decernis database.
i=1
Food fraud Food fraud types (sub- Code % Example
and for A ∈ U, we have types category)
∑ Replacement Fraudulent Labelling FLC 22.6 Selling cheddar
U{A} P (U, e) Claims cheese falsely
P(A|e) = (3)
P(e) labelled as ‘kosher’
(Incident ID:
The BN model for this study will comprise: 279451)
Dilution or Substitution DSAS 22.6 Contamination of
(1) A set of seven random variables (or nodes) of food fraud type, with an Alternate milk with harmful
Substance chemicals, powder,
country of origin, country of detection, food fraud vulnerability
urea, and polluted
factors, the weight of evidence, product types, and year. water (incident ID:
(2) A set of mutual states for each random variable (e.g. for the 20697)
variable food fraud type, the states are replacement, addition, Other/Dilution or ODS 17.2 Substitution of duck
and removal). Substitution with an meat with beef when
alternate ingredient tucking into self-
(3) A set of directed edges between the variables and an associated service at the
conditional probability for each of them. restaurant (Incident
ID: 116586)
The links between the variables in the BN are based on family re­ Misrepresentation of MAO 16.6 Selling catfish as
Animal Origin Atlantic wolfish to a
lationships. In this paper, food fraud vulnerability factor(s) is supposed
foreign market
to be the parent of food fraud type (and sub-type), year, country of (incident ID:
origin, country of detection, the weight of evidence, and product type. 1871079)
Relationships between food fraud type and other variables are drawn Use of Non-declared, UNB 8.1 Mixing hazardous
from the literature (see Bouzembrak & Marvin 2016; Marvin et al., unapproved, or Banned bleach chemical to
biocides the frozen meat
2016). For more details, an example of Bayesian Network calculations is products to make
presented in Appendix A (see Fig. A.1). them look fresh
The Bayesian Network model enables this research to learn patterns (Incident ID: 65650)
of relationship between the records attribute features (e.g. country of Misrepresentation of MGO 5 Relabelling hard to
Geographic Origin sell Mexican shrimp
origin, year, adulterants) from the training dataset and then apply these
as wild-caught
patterns to unknown instances (testing dataset) for food fraud vulnera­ American
bility assessment. Two Bayesian network models of Tree Augmented crustaceans’
Naïve (TAN) and Markov Blanket were selected as main classifiers for (Incident ID: 24799)
this research. The TAN model was selected out of the classical Naïve Misrepresentation of MVO 2.8 Blending original
Varietal Origin wine with wines of
Bayes because it allows each predictor (e.g. country of origin) to link
inferior quality
with another predictor (e.g. type of food products) in addition to the (Incident ID: 13398)
target variable (food fraud vulnerability factor). The Markov Blanket Formulation of an FAP 1.7 Serving dodgy drinks
model was selected to identify a subset of input variables (nodes) by Artificial Product as house spirits in
through the use of bars and pubs
reducing features with little predictive ability. The two TAN and Markov
multiple adulterants (Incident ID:
blanket networks were used to compare the performance of these and techniques 1044355)
Bayesian Network models in assessing food fraud vulnerability factors. Misrepresentation of MNC 0.5 Baby formula with
To construct the BN model for this paper, we will use 80 % of the inci­ Nutritional Content hazardously low
dent records from the Decernis food fraud database (as a training nutritional value
(Incident ID: 74149)
dataset), the machine learning technique ‘Expectation Maximisation
Algorithm’ and the SPSS Modeler 18.2 software. Addition Artificial Enhancement AEP 3.3 High levels of
of apparent Protein melamine in baby
content formula products
2.4. BN validation that might be from
animal feed (Incident
ID: 24574)
There are two ways to validate the constructed BN for this study. Artificial Enhancement AEC 1.9 Making fake beef
First, researchers will use approximately 20 % of the incident records of perceived quality made from pork and
(similar to the study by Marvin et al., 2016) in the Decernis food fraud with Colour additives treated with
additives such as
database to assess food fraud vulnerability factors. All variables (except
paraffin wax
food fraud vulnerability factors) will be used as input in the constructed (Incident ID: 17392)
BN model to assess food fraud vulnerability factors identified by the Artificial Enhancement AEO 1.6 Adulterating wines
Barrier Analysis Technique. Second, the study will compare the BN of Organoleptic by adding wood
model’s result in previous studies. For example, the result of the BN qualities chips for flavour
(Incident ID:
analysis for products will be compared with the research by Tähkäpää
1024237)
et al. (2015) and Bouzembrak and Marvin (2016).
Removal Removal of authentic RMVL 0 N/A
constituents
3. Results

3.1. Recorded food fraud incidents
All incidents recorded in the Decernis food fraud database were
extracted from the period 2000 to 2018 (accessed July 10, 2018). In­
cidents reported in the Decernis food fraud database linked to food fraud
were predominantly meat and poultry products (16.3%), dairy products
(16.1%), alcoholic beverage products (15.8 %), and seafood and seafood
products (12.4%). As these four product types cover more than 50 % of
the database related to incident records, this study focuses on these top
four product types as the sample for this paper (n = 580). The food fraud

3
F. Rezazade et al. Food Control 131 (2022) 108440

Table 3 Table 4
Fraud incidents’ country of origin and country of detection for all product types Results of the barrier analysis technique.
(n = 580). Food fraud Dimensions Description Source
Product types % Top countries of origin Top countries of detection vulnerability

Seafood 22.2 USA (27.13%), Taiwan USA (%34.1), UK (%11.6), Opportunity Detection of Identifying seafood BRC,
(20.08 %), China (12.4 %), Taiwan (%10.9), China (% adulterants species through Everstine,
UK (12.4%), Japan (9.3 %) 9,3), Japan (%6.2). requires advanced DNA analysis or Spink and
Meat and 29.7 China (21.6%), UK China (%22.8), UK (% laboratory analysis confirming the Kennedy
Poultry (19.13%), USA (10.49%), 20.4), USA (%9.3), South presence of (2013),
South Africa (6.79%). Africa (%6.8). methanol in Hoecht and
Alcoholic 27.8 China (13.67%), UK UK (%15.5), China (%13), alcoholic drinks Trott (2014), (
Beverages (12.42%), India (9.94 %), India (%9.9), Italy (%6.8) The physical form Ease of alteration by Decernis,)
Italy (9.32%). of products mixing raw Moyer,
Dairy 27.8 India (48.44%), China India (%48.4), China (% materials with low- DeVries and
(12.5%), Pakistan (6.25%), 14.8), USA (%7), Pakistan quality ingredients Spink (2017),
USA (7.3 %) (%6.3) or due to the nature Manning and
of the product (i.e. Soon (2016),
beef mince, ground Reilly (2018),
types were categorised in the Decernis food fraud database into three spices) Rezazade,
Availability of When the Summers and
distinct categories of addition, replacement and removal, as stated in
knowledge and knowledge/ Ong (2020),
Table 2. technology to technology required SSAFE (2017),
The majority of seafood fraud incidents were predominantly origi­ adulterate food for adulteration is Silvis et al.
nated and detected in the USA, UK, and Taiwan (see Table 3). Country of products generally available.’ (2017),Ting
Transparency Long and complex and Tsang
origin and country of detection linked to meat and poultry products and
supply chain supply chain due to (2014), Van
alcoholic beverages were predominantly China and the UK. India was globalisation Ruth, Huisman
identified as the most prominent country of origin and detection (48.4 Opportunity in Fraudsters have and Luning
%) for dairy products. Time and space legitimate access to (2017)
The majority of cited food fraud incident types were dilution/sub­ the location of
production and
stitution with alternative ingredient (23.8 %), dilution or substitution
processing lines
with an alternate substance (22.6 %), fraudulent labelling claims (22.6
Motivation Corruption level of High levels of
%), and misrepresentation of animal origin (16.6 %). The most cited
the country corruption in a
types of adulterants in the 580 cases of food fraud were the addition of country increase the
chemicals (33.3%) (i.e. to extend the shelf life of food products) and risk of fraud.
counterfeit type of adulterant (32.2%) (i.e. false declaration of organic/ The economic Financial strains
halal food product). health of the due to high
business (country) turnover, Culture,
the price spike, high
level of
3.2. Food fraud vulnerability factors competition, and
price differences
Supply and Pricing Economic pressures
The Barrier Analysis technique results revealed that the opportunity
due to supply
factor accounts for 32.7% of all incidents, the motivation factor accounts shortages
for 6.7% of all incidents, and the countermeasure factor accounts for
Countermeasure Requires When combating
46% of all incidents. The Unknown food fraud vulnerability factor was coordination fraud within an
related to incidents where there was no information provided (either between law international level
unavailable or in other languages). Table 4 shows a description of these enforcement (in all Europe, for
factors with their underlying dimensions. The result of the Barrier agencies example), or when
there is a need in
Analysis technique (and inclusion of these factors and their sub-factors)
border regions to
for each incident was validated/confirmed by the literature, food stop adulterated
regulation reports, and expert judgment. food reaching
customers.
Extensiveness of Using reactive track
Traceability and tracing systems
3.3. Bayesian Network modelling that can be easily
imitated, lack of
A Bayesian Network model was proposed using the records from the packaging and
Decernis food fraud database and the Barrier Analysis Technique results. labelling system
Lack of law No punishment,
Table 5 presents the input variables used to construct the BN model to
enforcement although a violation
assess food fraud vulnerability factors (the target variable), measure­ of laws are apparent
ment types, and the input variable source. The input variables shown Food safety Detectability is
with binary measurement categorisation (the type of fraudulent in­ based on food safety
sampling and food
cidents, country of detection, and type of adulterant) were found to
quality, not Food
occur in the database in multiple records. Thus it was essential to apply a Fraud Vulnerability
binary code to each category within these variables to avoid multiple assessment, lack of
entries recorded for a single cell. This resulted in the identification of definition of fraud
128 input variables for the Bayesian network. in law
Whistleblowing Lack of a well-
Fig. 1 shows the Bayesian Model Stream constructed by IBM SPSS
system designed
Modeler version 18.2. The steps used to create this Stream were: whistleblowing
(continued on next page)
1. Import the external source file as a ‘Source’ node.

4
F. Rezazade et al. Food Control 131 (2022) 108440

Table 4 (continued )
Food fraud Dimensions Description Source
vulnerability
system (e.g. former
employees that
were sacked from
companies are
reporting
fraudulent
activities)
2. Apply the ‘Partition’ node to divide the dataset into a training (80 %)
and a testing set (20 %) similar to studies by Bouzembrak and Marvin
(2016) and Marvin et al. (2016).
3. Apply the ‘Type’ node to set the variable’s role, set the food fraud
vulnerability field to ‘Target’, and the remainder of the variables to
‘Input Variables’ (except record ID, which was set to Record ID).
4. Exclude cases where the target had a null value using a ‘Select’ node.
5. Create two Bayesian Network models of Tree Augmented Naïve
(TAN) and Markov Chain Model - with Maximum Likelihood
parameter learning method. This results in two model nuggets being
generated.
When executing the TAN and Markov models, two corresponding
Model Nuggets (gold diamond-shaped icon) are placed on the Stream.
The Model Nuggets shows details about the importance of the input
variables in building the model through the predictor importance
(similar to sensitivity analysis) chart. These two models were compared
by applying a ‘Filter’ node and creating a Gains chart and output anal­
ysis node. The Gains chart (Fig. 3) illustrates the result that the TAN gain
(blue line) is far superior to the Markov model gain (red line) in
assessing food fraud vulnerability factor. The output analysis provides
an opportunity to view predictions (data) and check the model’s accu­
racy for assessing food fraud vulnerability factors. TAN model correctly
identified 86.09 % of the food fraud vulnerability outcomes than 49.57
% of the Markov model (see Fig. 2).
3.3.1. Predictor importance
All input variables were selected as influencing variables, and the
Bayesian model was tested to determine the most important influencing
variables of food fraud vulnerability factors. By opening a TAN model
nugget, various details about the importance of input variables in pre­
dicting food fraud vulnerability factors can be seen through the pre­
dictor importance chart.
The SPSS Modeler has the feature of predictor importance (through
the model nuggets) that is derived from the sensitivity analysis of the
model. Fig. 4 shows the network between food fraud vulnerability fac­
tors and all the important variables, and Fig. 5 shows the predictor
importance of the target node (food fraud vulnerability factors). The
country of origin (76 %) variable was the most relevant when assessing
food fraud vulnerability from the predictor importance analysis.
The next variables important in this assessment were (in order): the
type of food product (9 %); Counterfeiting (as a type of adulterant) (9
%); and China (as a detecting country) (6 %) (see Table 6). The predictor
importance analysis for all variables known to play a role in assessing
vulnerability to food fraud was conducted and the results provided in
the form of Conditional Probability Tables (CPTs).
Appendix B provides CPTs for the country of origin (developed by
TAN model Nuggets) (see Table B.1). The result of CPTs shows three
combinations of conditions that help understand the factors that
contribute to a high probability of vulnerability for food fraud when we
know the country of origin. These combinations are:

Table 5
Dependant (food fraud vulnerability) and independent variables.
Variable Name Data type/ Description Reference
Measurement

Record ID Categorical/ Record ID Decernis, Moore, Spink and Lipp (2012),

Nominal Marvin et al. (2016)
Type Categorical/ Food product types (seafood, meat and poultry, alcoholic beverages, dairy products)
Nominal
Food fraud type Categorical/ Replacement, Addition, Removal
Binary/Flag
Type of Categorical/ Additive (i.e. colour.), Chemicals (i.e. bleach), Expired (data of production),
adulterants Binary/Flag Counterfeit (i.e. fake halal certificate), Species (i.e. false declaration of species of
food products)
Weight of Categorical/ Low, Medium, High
Evidence Ordinal
Country_ O Categorical/ Country of Origin
Nominal
Country_ D Categorical/ Country of Detection
Binary/Flag
Year Categorical/ Food fraud incident year from 2000 to 2018
Binary/Flag

Food fraud Categorical/ Food fraud vulnerability factors (Opportunity, Motivation, Countermeasures, Barrier Analysis Technique, literature,
vulnerability Nominal Unknown) government reports and expert judgement (See
Table 3)

Fig. 1. Spss modeler stream.

5
F. Rezazade et al.
Fig. 2. Analysis results for Tan and Markov Chain models.
1. Alcoholic Beverages are most vulnerable to food fraud motivated by
the desire for economic gain or as a result of cultural influences
(motivation), when those beverages originate from India.
2. Dairy products are most vulnerable to food fraud as a result of
complex supply chains (opportunity) when they originate from
India; and
3. Seafood products are most vulnerable to food fraud as a result of lack
of advanced laboratory testing for detection of fraud (opportunity)
when they originate from the USA (see highlighted cells in appendix
B).
3.4. Bayesian Network validation
There were two ways to verify the validity of the TAN Network
Model developed in this research. First, the data was divided into a
Fig. 3. Gain chart result for TAN and Markov Chain models.
6
Food Control 131 (2022) 108440
training (80 %) and a testing (20 %) dataset by the Partition Node (see
Fig. 1). The TAN model constructed based on 80 % (465 cases) of the
total data collected from the Decernis food fraud database was used to
assess vulnerability to food fraud in 20 % (115 cases) of the testing
subset. The testing sub-set was selected randomly from the whole
dataset. All variables (except food fraud vulnerability factors) were
selected as input parameters in the TAN model to assess food fraud
vulnerability. The TAN model can assess vulnerability to food fraud in
86 % of the cases with a probability of more than 0.5 (see Appendix C,
Table C.1 in the Appendices).
Second, the output of the TAN model was compared to the results
previously addressed by other researchers. Marvin et al. (2016) devel­
oped a Bayesian Model to predict food fraud type in 91.5 % of cases.
These authors reported a similar result for predictor importance. They
reported that the country of origin was the most critical factor (61 %) in
determining the type of food fraud. These authors used the BN model
based on other databases (RASFF and EMA) to predict the type of food
fraud and have used several food safety drivers as predictors.
The Bayesian modelling approach adopted in this study used real-
data from internationally reported food fraud cases in the Decernis
food fraud database to assess vulnerability to food fraud for different
food types from various countries considering the three Routine Activity
Theory classifications (Opportunity, Motivation and Countermeasures).
The holistic approach based on the BN model is shown in Fig. 6. The
paper identified four key steps to be followed to achieve this outcome:
• Step 1: review previous food fraud cases recorded in global food
fraud databases (Decernis food fraud database).
• Step 2: identify food fraud vulnerability factors using Barrier Anal­
ysis techniques and organise these factors into the classifications, as
indicated in the Routine Activity Theory framework (opportunity,
motivation, and countermeasures).
• Step 3: develop a BN model based on the extracted factors and other
variables addressed in this paper; and
• Step 4: determine proactive measures to combat the specified food
fraud vulnerability areas.
F. Rezazade et al.
Fig. 4. Tan model.
4. Discussion
The food fraud literature has repeatedly called for an empirically
tested, holistic approach to combating and preventing food fraud
through the identification of areas of vulnerability (Lord et al., 2017;
Silvis et al., 2017; Spink & Moyer 2011; Spink et al., 2017; Van Ruth
et al., 2017). This call from the academy is also mirrored by food stan­
dards agencies, where for example, compliance policies from organisa­
tions such as the Global Food Safety Initiative (GFSI) (2018), now also
mandate the inclusion of vulnerability factors in auditing assessments.
The Gloabal pandemic of 2020 has also highlighted the importance of
food security (World Bank, 2021) with the World Health Organisation
(WHO) now also recommending that food fraud vulnerability assess­
ment be included in prevention strategies for food production industries
(WHO, 2020).
Despite the recognised importance of vulnerability assessment in the
mitigation and prevention of food fraud, there remain discrepancies
between the approaches used by food standard agencies and those
proposed by academic researchers to identify and manage food fraud
vulnerability. These discrepancies have resulted in an inconsistency in
approaches used to review food fraud incidents and a consequential lack
of objectivity in identifying and assessing vulnerability factors. If
stakeholders from food safety agencies and those involved in food
Fig. 5. Predictor importance.
7
Food Control 131 (2022) 108440
production are to comply with the standards required by GFSI (2018)
and WHO (2020), these discrepancies need to be addressed.
This paper takes up the call to develop an empirically tested, holistic
approach to combating and preventing food fraud through the objective
identification and assessment of areas of food fraud vulnerability. The
model used in this study, has addressed the discrepancies between
practioners and academic researchers in this area by adopting a
harmonised framework that brings together both approaches. Scholars
have long advocated for the use of data from historical food fraud in­
cidents as an ideal way to gather relevant information about vulnera­
bilities and the factors that influence them. To date however, there has
not been a useful way to review this volume of data informed by pre-
defined influencing variables for food fraud, drawn from the experi­
ence of food standard agencies.
This approach not only provides a more objective analysis of the
predictors of vulnerability to food fraud but it also provides recom­
mendations for the best deterrent strategies for each type of vulnera­
bility. The model proposed in this study was based on a Bayesian
network analysis of the Decernis database of 580 food fraud incidents
and included all known influencing variables defined by SSAFE (2017)
and those proposed by academic scholars. The Bayesian model was able
to assess food fraud vulnerability with an accuracy of 86 %. The model
itself could be adapted for future use by border protection and food
safety agencies to improve the ability for simultaneous identification
and prevention of food fraud vulnerability in real-time.
The results of this study indicated that, on average, the collective
contributions of the variables Country of Origin and Product Type, were
the most influential in assessing vulnerability to food fraud. This finding
is especially important for the modern network of the interconnected
international supply chains for food production, as it highlights the need
for international agencies to be transparent and cooperative in their
policies for food fraud detection and prevention across the entire supply
chain.
While applying these results to detection, mitigation, and counter­
measure strategies may be criticized from the perspective of discrimi­
nation (particularly in relation to country of origin), the findings align
with those from other studies that have examined predictors of food
Table 6
Predictor analysis results.
Nodes Description Importance
FLC Fraudulent Labelling Claim 0.0042
China Country of Detection 0.0558
CNTRFT Counterfeit (adulterant type) 0.0861
Type Food product types 0.0945
Country_O Country of origin 0.7595
F. Rezazade et al.
Fig. 6. A holistic approach to food fraud vulnerability assessment.
fraud (Marvin et al., 2016; Berman 2008). Indeed, Berman (2008) spe­
cifically cautioned exporters to carefully control and manage their food
supply chains and be wary of counterfeiting issues when engaging in the
growing China market (Berman, 2008).
In addition to Country of Orign and Product Type, the predictor
importance analysis applied in this study, also highlighted that:
detecting the country where the fraud occurs; the types of adulterants
employed; and the incident types greatly improved the predictive power
of the Bayesian model (Fig. 5). This finding was not expected as the
interrelationships of these variables had not previously been suggested
in the extant literature.
5. Conclusion
This study is the first to apply a Bayesian modelling approach to
interrogate live data from global food fraud incidents (using the
Decernis database) along with secondary data about the variables
known to influence vulnerability to food fraud. This holistic approach
analysed large volumes of data to provide useful insights for both
practioners and scholars to better understand the areas of vulnerability
to mitigate and defend against food fraud. The model was also the first to
take a vulnerability identification approach in the analysis of food fraud
data from multiple food product types and countries of origin, thus
increasing the objectivity and applicability of the results.
The results confirmed that country of origin, product type, detecting
country, and types of adulterants were the most critical variables in the
assessment of food fraud vulnerability. The CPTs provided empirical
evidence that combining factors contributes to a higher probability of
predicting vulnerability to food fraud than considering these factors
independently. Applying this finding to food import data will allow
border authorities to better determine the likelihood of vulnerability to
food fraud and the best countermeasures to reduce vulnerability and/or
inform decisions to reject the product from entry altogether.
Adapting and extending the model constructed for this paper, to
include new variables based on expert judgement, other data sources
8
Food Control 131 (2022) 108440
and improved mobility will provide additional predictive power for key
decision-makers in real-time. Similarly, adding the level of hazard po­
tential as an additional input variable for each food fraud incident would
add new insights to our understanding of how these variables interact.
Currently, the model requires the skills of a person proficient in pro­
gramming and mathematics to manipulate and interpret the data for
application by practitioners. Developing an interface that would remove
this requirement and allow data manipulation at a more simplistic level
will increase the usefulness and accessibility of the model and the
findings and form the basis for future research.
Food security is a growing 21st-century problem, highlighting the
need for coordinated international food fraud prevention and mitigation
systems in global food supply chains. The results presented in this paper
affirm the need for coordination and transparency between interna­
tional stakeholders to ensure a strong sustainable food supply chains for
recovering post covid-19 global economies.
CRediT authorship contribution statement
Faeze Rezazade: Conceptualization, Methodology, Writing – orig­
inal draft, Writing – review & editing. Jane Summers: Supervision,
Conceptualization, Writing – review & editing. Derek Ong Lai Teik:
Supervision, Writing – review & editing.
Declaration of competing interest
The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influence
the work reported in this paper.
Acknowledgement
This work was supported by the University of Southern Queensland
Postgraduate Research Scholarship (USQ PGRS).
F. Rezazade et al. Food Control 131 (2022) 108440

Appendix A

Example of Bayesian Network Calculations

Figure A1 shows an example of a BN with three nodes (random variables) of Food Fraud Vulnerability (FFV) factors, Food Fraud Type (FFT), and
Year (Y). Variable FFV factors have three states of Opportunity (O), Motivation (M), and (Countermeasures), variable FFT has three states of
Replacement (Rep), Addition (Add), and Removal (Rem), and finally, variable Year has four states of Y1 (2002–2006), Y2 (2006–2010), Y3
(2010–2014), and Y4 (2014–2018). Based on the literature, FFV and FFT are parents of Year. Variable Year (Y) with its parents FFV and FFT, form a
conditional probability distribution, P(Year|FFV, FFT) or Conditional Probability Table (CPT). CPTs and Unconditional Probability Tables (UPTs) are
shown in Figure A1. The UPTs are given to a variable (e.g. FFV and FFT) when there is no parent (Jensen & Nielsen 2007). In this example, we
supposed that FFV and FFT are independent. UPTs and CPT can be obtained through training the data (Marvin et al., 2016, p. 469). In the case that
there are no known UPTs, n1 probability can be assigned, where n is ‘the number of states of the parameters’ (Marvin et al., 2016, p. 469).
Child variable (Y) is conditionally dependent on parent variables’ states (FFV and FFT). The CPT for the child variable (Y) is a table of all possible
stated combinations of the parent variables in Figure A1. The first probability in the CPT in Figure A1 can be described as: given that FFV is (O), and
FFT is (Rep), the probability that (Y1) will be equal to them is the conditional probability ​ P(Y = Y1| ​ FFV ​ = 0, ​ FFT = Rep).

Fig. A.1. Example of BN calculation

Appendix B
Table B.1
CPT for the country of origin

Parents Conditional Probability Table- Country of Origin

Type FFV India USA

Seafood Opportunity 0.07 0.5

Motivation 0 0
Countermeasure 0.02 0.16
Unknown 0.09 0.18

Meat and Poultry Opportunity 0 0.03

Motivation 0 0.17
Countermeasure 0.02 0.12
Unknown 0 0.15

Alcoholic Beverages Opportunity 0.05 0.03

Motivation 0.58 0.08
Countermeasure 0.05 0.05
Unknown 0.09 0

Dairy Opportunity 0.58 0.09

Motivation 0.8 0
Countermeasure 0.27 0.08
Unknown 0.32 0.04

9
F. Rezazade et al. Food Control 131 (2022) 108440

Appendix C
Table C.1
Assessment Accuracy

ID FFV TAN $BP-FFV

1528051 3 3 1
913178 1 1 1
289925 2 2 0.99758567
289919 2 2 0.99758567
197572 1 1 0.989840435
112951 3 3 0.876018242
90337 3 3 0.977842454
68861 3 3 0.992059917
62896 3 3 0.999520531
58076 3 3 0.999975065
56019 1 1 0.998240596
51808 3 3 0.999520531
51462 3 3 1
17419 4 4 1
17213 4 4 1
17141 3 3 0.741353187
16780 1 1 0.999999048
16778 1 1 0.999999048
16774 1 1 0.999999048
16772 1 1 0.999999048
16086 3 3 0.945414977
13319 1 1 0.607666094
13296 1 1 0.607666094
788492 3 3 1
390535 3 3 0.936464035
121420 3 3 1
22770 3 3 1
22731 3 3 0.918745204
213060 3 3 1
122912 3 3 0.906289431
118699 1 1 1
94320 2 2 1
93111 2 2 1
69794 1 1 0.997960086
65654 3 3 0.987642311
65650 3 3 0.987642311
63306 3 3 1
54159 3 3 0.707213208
51537 3 3 1
34113 3 3 1
27179 4 4 1
26773 3 3 0.998111932
25961 1 1 0.945788698
25475 3 3 0.973100091
23313 3 3 0.999237774
22293 3 3 1
22166 1 1 0.960104719
22161 1 1 0.960104719
20303 4 4 1
18778 3 3 1
17361 3 3 1
17269 3 3 0.98865135
17267 3 3 0.98865135
16499 3 3 1
15261 3 3 1
13557 1 1 0.885504533
13361 2 2 1
1629449 2 2 1
1044095 3 3 0.926555131
828615 4 4 1
191531 1 1 1
269902 3 3 1
244848 4 4 0.511281743
91946 1 1 0.987555274
66589 1 1 1
34363 3 3 1
34359 3 3 1
34201 3 3 0.877191916
34085 3 3 1
30170 4 4 1
25575 2 2 0.985500917
25463 2 2 0.999989005
23185 3 3 1
(continued on next page)

10
F. Rezazade et al. Food Control 131 (2022) 108440

Table C.1 (continued )

ID FFV TAN $BP-FFV

22406 3 3 0.94880778
13747 4 4 1
13724 4 4 1
13417 2 2 0.999509216
13398 3 3 1
13260 1 1 0.811741271
1596726 1 1 0.909537089
631001 4 4 1
279451 3 3 1
461064 1 1 0.520506488
61718 4 4 1
53875 2 2 1
53256 1 1 0.970646895
50469 2 2 0.991298833
50457 2 2 0.997488168
33748 1 1 0.999666777
31106 1 1 0.900078711
29797 1 1 0.7916972
23594 4 4 1
22960 1 1 0.855686946
22570 1 1 0.918031127
17207 4 4 1
16450 3 3 0.968345746
13282 3 3 0.498705735
13224 1 1 0.940245524

References Moyer, D. C., DeVries, J. W., & Spink, J. (2017). The economics of a food fraud incident-
case studies and examples including melamine in wheat gluten. Food Control, 71.
https://doi.org/10.1016/j.foodcont.2016.07.015
Berman, B. (2008). Strategy to detect and reduce counterfeiting activity. Business
Rasff. Rapid Alert System for Food and Feed database. from European Commision
Horizons, 51(3), 191–199. https://doi.org/10.1016/j.bushor.2008.01.002
https://ec.europa.eu/food/safety/rasff/portal_en.
Bouzembrak, Y., & Marvin, H. J. P. (2016). Prediction of food fraud type using data from
Reilly, A. (2018). Overview of food fraud in the fisheries sector (C1165). FAO Retrieved
Rapid Alert System for Food and Feed (RASFF) and Bayesian network modelling.
from http://www.fao.org/3/I8791EN/i8791en.pdf.
Food Control, 61, 180–187. https://doi.org/10.1016/j.foodcont.2015.09.026
Rezazade, F. (2020). The assessment of factors contributing to food fraud vulnerability: A
Bouzembrak, Y., Steen, B., Neslo, R., Linge, J., Mojtahed, V., Marvin, H. J. P., et al.
bayesian modelling approach on a food fraud database. PhD dissertation. Toowoomba,
(2018). Development of food fraud media monitoring system based on text mining.
Australia: University of Southern Queensland http://eprints.usq.edu.au/id/eprint
Food Control, 93, 283–296. https://doi.org/10.1016/j.foodcont.2018.06.003
/39501.
BRC. Understanding vulnerability assessment. Retrieved from https://www.siroccocons
Silvis, I. C. J., Van Ruth, S. M., Van der Fels- Klerx, H. J., & Luning, P. A. (2017).
ulting.com/wp-content/uploads/2017/12/UNDERSTANDING-VULNERAB
Assessment of food fraud vulnerability in the spices chain: An explorative study.
ILITY-ASSESSMENT-BRC-2017.pdf.
Food Control, 81. https://doi.org/10.1016/j.foodcont.2017.05.019
Decernis. Food Fraud Database. Retrieved June 2018 https://decernis.com/solutions
Soon, J. M. (2020). Application of bayesian network modelling to predict food fraud
/food-fraud-database/.
products from China. Food Control, 114. https://doi.org/10.1016/j.
Decernis. Protect Your Brand – Combat Food Fraud. Retrieved from https://decernis.
foodcont.2020.107232, 107232.
com/solutions/food-fraud-database/.
Spink, J., Elliott, C., Dean, M., & Speier-Pero, C. (2019a). Food fraud data collection
Everstine, K., Spink, J., & Kennedy, S. (2013). Economically motivated adulteration
needs survey. Native Plants Journal Food Science, 3, 8. https://doi.org/10.1038/
(EMA) of food: Common characteristics of EMA incidents. Journal of Food Protection,
s41538-019-0036-x
76(4), 723–735. https://doi.org/10.4315/0362-028X.JFP-12-399
Spink, J., Fortin, N. D., Moyer, D. C., Miao, H., & Wu, Y. (2016). Food fraud prevention:
FDA. (2009). An overview of the Carver Plus Shock method for food sector vulnerability
Policy, strategy, and decision-making – implementation steps for a government
assessments. https://www.fda.gov/files/food/published/CARVER—SHOCK-PRIME
agency or industry. Chimia, 70(5), 320–328. https://doi.org/10.2533/
R-PDF.pdf.
chimia.2016.320
GFSI. (2018). Tackling food fraud through food safety management systems. Retrieved
Spink, J., & Moyer, D. C. (2011). Defining the public health threat of food fraud. Journal
from https://mygfsi.com/wp-content/uploads/2019/09/Food-Fraud-GFSI-Technica
of Food Science, 76(9), R157–R163. https://doi.org/10.1111/j.1750-
l-Document.pdf.
3841.2011.02417.x
Hoecht, A., & Trott, P. (2014). How should firms deal with counterfeiting? International
Spink, J., & Moyer, D. C. (2013). Understanding and combating food fraud food Technology
Journal of Emerging Markets, 9(1), 98–119. https://doi.org/10.1108/ijoem-02-2011-
magazine (Vol. 67).
0014
Spink, J., Moyer, D. C., & Speier-Pero, C. B. (2016). Introducing the food fraud initial
Hossain, M., & Muromachi, Y. (2013). A real-time crash prediction model for the ramp
screening model (FFIS). Food Control, 69. https://doi.org/10.1016/j.
vicinities of urban expressways. IATSS Research, 37(1), 68–79. https://doi.org/
foodcont.2016.03.016
10.1016/j.iatssr.2013.05.001
Spink, J., Ortega, D. L., Chen, C., & Wu, F. (2017). Food fraud prevention shifts the food
Jensen, F. V., & Nielsen, T. D. (2007). Bayesian networks and decision graphs. USA:
risk focus to vulnerability. Trends in Food Science & Technology, 62, 215–220. https://
Springer.
doi.org/10.1016/j.tifs.2017.02.012
Lord, N., Flores, E., Cecilia, J., & Spencer, J. (2017). The dynamics of food fraud: The
Spink, J., Vincent Hegarty, P., Fortin, N. D., Elliott, C. T., & Moyer, D. C. (2019b). The
interactions between criminal opportunity and market (dys)functionality in
application of public policy theory to the emerging food fraud risk: Next steps.
legitimate business. Criminology and Criminal Justice, 17(5), 605–623. https://doi.
Trends in Food Science & Technology, 85, 116–128. https://doi.org/10.1016/j.
org/10.1177/1748895816684539
tifs.2019.01.002
Manning, L., & Soon, J. M. (2016). Food safety, food fraud, and food defense: A fast
SSAFE. (2017). Food fraud vulnerability assessment tool. Retrieved from https://www.
evolving literature. Journal of Food Science, 81. https://doi.org/10.1111/1750-
ssafe-food.org/our-projects/.
3841.13256
Tähkäpää, S., Maijala, R., Korkeala, H., & Nevas, M. (2015). Patterns of food frauds and
Marvin, H. J. P., Bouzembrak, Y., Janssen, E. M., Van der Fels- Klerx, H. J., van
adulterations reported in the EU rapid alert system for food and feed and in Finland.
Asselt, E. D., & Klete, G. A. (2016). A holistic approach to food safety risks: Food
Food Control, 47, 175–184. https://doi.org/10.1016/j.foodcont.2014.07.007
fraud as an example. Food Research International, 89. https://doi.org/10.1016/j.
Tien, I., & Der Kiureghian, A. (2016). Algorithms for Bayesian network modeling and
foodres.2016.08.028
reliability assessment of infrastructure systems. Reliability Engineering & System
Moore, J. C., Spink, J., & Lipp, M. (2012). Development and application of a database of
Safety, 156, 134–147. https://doi.org/10.1016/j.ress.2016.07.022
food ingredient fraud and economically motivated adulteration from 1980 to 2010.
Ting, S. L., & Tsang, A. H. C. (2014). Using social network analysis to combat
Journal of Food Science, 77(4), R118–R126. https://doi.org/10.1111/j.1750-
counterfeiting. International Journal of Production Research, 52(15), 4456–4468.
3841.2012.02657.x
https://doi.org/10.1080/00207543.2013.861947

11
F. Rezazade et al.
Van Ruth, S. M., Huisman, W., & Luning, P. A. (2017). Food fraud vulnerability and its
key factors. Trends in Food Science & Technology, 67, 70–75. https://doi.org/
10.1016/j.tifs.2017.06.017
Van Ruth, S. M., Luning, P. A., Silvis, I. C. J., Yang, Y., & Huisman, W. (2018). Differences
in fraud vulnerability in various food supply chains and their tiers. Food Control, 84,
375–381. https://doi.org/10.1016/j.foodcont.2017.08.020
Wiśniewska, M. (2017). CARVER+Shock method and its application in a catering
company in Poland. British Food Journal, 119(12), 2610–2629. https://doi.org/
10.1108/BFJ-11-2016-0554
World Bank. (2021). Food security and COVID19. Retrieved from https://www.worldban
k.org/en/topic/agriculture/brief/food-security-and-covid-19.
12
Food Control 131 (2022) 108440
World Health Organisation (Who). (2020). COVID-19 and food safety: Guidance for
competent authorities responsible for national food safety control systems. Retrieved
from https://www.who.int/publications/i/item/covid-19-and-food-safety-guida
nce-for-competent-authorities-responsible-for-national-food-safety-control-systems.
Yan, J., Erasmus, S. W., Aguilera Toro, M., Huang, H., & van Ruth, S. M. (2020). Food
fraud: Assessing fraud vulnerability in the extra virgin olive oil supply chain. Food
Control, 111. https://doi.org/10.1016/j.foodcont.2019.107081
Zhang, W., & Xue, J. (2016). Economically motivated food fraud and adulteration in
China: An analysis based on 1553 media reports. Food Control, 67, 192–198.