A Machine Learning Approach for Privacy-preservation in
E-business Applications
Fatemeh Amiri
1,2
, Gerald Quirchmayr
1,2
, Peter Kieseberg
3
1
University of Vienna, Vienna, Department of Computer Science, Austria
2
SBA Research Institue, Vienna, Austria
3
St. Poelten University of Applied Sciences, St. Poelten, Austria
Keywords: Privacy-preserving, E-business, Big Data, Data Mining, Machine Learning.
Abstract: This paper aims at identifying and presenting useful solutions to close the privacy gaps in some definite data
mining tasks with three primary goals. The overarching aim is to keep efficiency and accuracy of data mining
tasks that handle the operations while trying to improve privacy. Specifically, we demonstrate that a machine
learning methodology is an appropriate choice to preserve privacy in big data. As core contribution we
propose a model consisting of several representative efficient methods for privacy-preserving computations
that can be used to support data mining. The planned outcomes and contributions of this paper will be a set
of improved methods for privacy-preserving soft-computing based clustering in distributed environments for
e-business applications. The proposed model demonstrates that soft computing methods can lead to novel
results not only to promote the privacy protection, but also for retaining performance and accuracy of regular
operations, especially in online business applications.
1 INTRODUCTION
Privacy-preservation in e-business that holds big data
is an essential topic of discussion and, requires
addressing several aspects (Verykios et al., 2004a).
Applying a privacy improvement task may bring side
effects to other running tasks like data mining
applications and administration programs.
Managing big data is itself an extensive field and
is made possible by involving a knowledge discovery
concept which is the process of extracting useful
hidden information from raw data (Fayyad et al.,
1996, Narwaria and Arya, 2016). Data mining, as the
core part of the knowledge discovery approach,
enables the practitioner to extract useful knowledge
from the data to better understand and serve their
customers to gain competitive advantages (Chen,
2006).
From the privacy-preserving point of view,
confidential and sensitive data should be protected at
the same time when running the data mining process
(Jahan et al., 2014). Therefore, the long-term aims in
this paper are specific attainments in privacy-
preserving data mining (PPDM) in the e-business
environment that can be achieved by a certain number
of steps.
In the first studies on PPDM, the primary
objective of privacy-preserving algorithms was only
protecting sensitive data at the time of publishing
(Westin, 1999). Lots of approaches are adopted to
solve this issue. Some of the proposed algorithms try
to hide the sensitive data and some others are
designed to modify and add extra noise in a manner
to protect raw data (Samet and Miri, 2012). Also, data
mining tasks are time-consuming.
So, if privacy preserving approaches reduce the
performance significantly, the final combined
algorithms will most likely not be useful. This issue
is even more problematic in distributed environments
than centralized models.
Hence, a strong privacy-preserving method
cannot present real practical achievements without
satisfying complexity constraints.
As Figure 1 shows, privacy of data should be
protected as vigorously as possible while at the same
time assuring an acceptable level of performance and
accuracy.
However, the accuracy of data mining methods
should be fixed and overheads must be as low as
possible. Thus, privacy-preservation is not the only
issue anymore. Accuracy and the performance of the
data mining tasks have the same necessity. In this
Amiri, F., Quirchmayr, G. and Kieseberg, P.
A Machine Learning Approach for Privacy-preservation in E-business Applications.
DOI: 10.5220/0006826304430452
In Proceedings of the 15th International Joint Conference on e-Business and Telecommunications (ICETE 2018) - Volume 2: SECRYPT, pages 443-452
ISBN: 978-989-758-319-3
Copyright © 2018 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
443

Scalability
parallelization
resilience
imprecision
Privacy
Protection
Accuracy
Effieciency
Main Objective
Other Factors
Secondary Objective
Figure 1: Relationship between PAE keywords (Privacy-
Accuary-Efficiency).
paper, equal attention will be paid to these three key
parameters - PAE (Privacy, Accuracy, Efficiency). It
should be noted that in e-business applications, some
other factors like those mentioned in Figure 1 come
into action and need to be taken into account by
design approaches. Lots of models and algorithms
were proposed in order to close privacy gaps.
However, privacy is not the only point, and other
factors need to be taken into account. To the best of
our knowledge, a method with such a description that
satisfies all PAE factors has not yet been proposed.
Solving this issue needs a new approach to reach
significant improvements. A considerable number of
scholars using conventional methods indicate the
requirement for a new technique.
Soft computing as a collection of various
methodologies could provide flexible knowledge
processing facilities to handle ambiguous real-world
problems in a synergistically manner (Zadeh, 1994).
The idea of using the word soft indicates the focus of
using methods that are not conventional or hard.
However, in many publications soft computing is also
used as synonym for machine learning too. Soft
computing can focus on achieving robustness,
tractability, and at the same time provide less
expensive solutions regarding time and space
complexity (Maimon and Rokach, 2007). As a result,
the challenge of developing acceptable solutions for
privacy-preserving data mining are well-addressed by
soft computing methods (Mitra et al., 2002) (Malik et
al., 2012).
Recently, privacy-preserving using soft
computing methods was discussed and some novel
approaches proposed by the authors of this paper
(Amiri and Quirchmayr, 2017). These methods yield
better results in all PAE aspects defined in Figure 1.
This paper will build on our previous work by
further exploring and describing privacy gaps in
PPDM and then propose some methods to close
privacy gaps in selected data mining tasks.
Regarding the problem of closing privacy gaps
and harmonizing some other factors, it is expected to
make a number of improvements that contribute to
advancing the state of the art in privacy-preservation
in data mining in distributed environments:
 A privacy-preserving model for soft computing-
based clustering. This model consists of several
methods for improving the PAE factors. Each
method uses a different soft computing approach
to find a reasonable level of performance for each
factor.
 The proposed algorithms bring a better level of
privacy. With implementing soft computing
methods, we expect to reach better results than
traditional methods of PPDM. The definition of a
privacy breach in this paper is: Disclosing not
only the sensitive information of individuals
without their consent, but also releasing
information in such a way that the individuals may
be linked with totally unrelated information even
with a very small probability.
 The accuracy of data mining task results. The
proposed algorithms shall not sacrifice accuracy
of data mining tasks. These methods protect the
privacy of data before the mining task and, keep
the utility and do not alter raw data.
 The efficiency of overall operations. As soft
computing methods are natively fast, we hope that
the computation and communication cost would
not significantly increase through the addition of
privacy-preserving components. Figure 2 shows
the overall plan.
As Figure 2 shows, SOM(Self Organizing Map)
clustering data mining tasks is used for data mining
purposes. This clustring could be horizontally or
vertically in distribution. The primary objective is
accuracy and efficiency of the final output of data
mining tasks with a high privacy level. To achieve
this purpose, we use a combination of soft computing
methods – depicted in the left box of the picture. As
discussed before, none of the unique techniques could
satisfy all PAE factors. So a hybrid model will be
used in the proposed algorithms. We apply a Neural
Network based clustering like SOM and try to
promote its PAE factors. According to the current
state of research, SOM clustering suffers from lack of
privacy, as well as a lack of some communication
aspects in the distributed interaction between clusters.
SECRYPT 2018 - International Conference on Security and Cryptography
444

Figure 2: The overall plan of proposed approach in this paper including the soft methods in the left box, selected data mining
task and expected results.
We try to close these gaps by implementing soft
computing methods like fuzzy sets and Gas(Genetic
Algorithms) in our proposed plans.
The core of this paper is organized as follows:
Related work on privacy-preservation in data mining
and soft computing is reviewed in Section 2. Section
3 discusses the definition of the problem with the
parameters involved. The applied methodology to
design the model and related methods is also covered
in this section. Model schemes for implementing the
new privacy-preserving mechanisms are described in
Section 4. In Section 5, we discuss the application of
our mechanisms to protect privacy when using
machine learning and also discuss expected results.
Finally, in Section 6, we summarize the major
conclusions of our research and outline future
research directions.
2 RELATED WORKS
Privacy-preserving data mining was introduced by
(Agrawal and Srikant, 2000). Related studies present
several useful methods with different views on the
PAE factors. Some scholars focus only on one kind
of database, like (Agrawal and Srikant, 2000) that
concentrate on statistical databases and promotions
brought by reconstructing the perturbation before
constructing the mining process. However, in e-
business applications, data transitions usually happen
in distributed environments. (Wu et al., 2007)
proposed a simplified taxonomy that divides the data
mining process into two centralized and distributed
environments. Despite its importance, the methods
presented are much less suited for distributed
environments than central servers. Another important
issue lies in the selection of the method for achieving
privacy (e.g., k-anonymity through generalization),
as this selection directly influences the PAE factors.
The capability of candidate methods to improve
privacy should not be the only factor in this selection.
Some other factors like accuracy and complexity
should be taken into account as well. However, this
point is rarely investigated by scholars. (Aggarwal
and Philip, 2008) used a general classification to
explore current research, including randomization
methods, the k-anonymity model, and distributed
privacy preservation. A peer contribution of their
study (that has not been investigated before) is paying
particular attention to the side effects of data mining
approaches that may influence the privacy of
individuals. They explore this side effect as
Downgrading Application Effectiveness. Although
they propose an effectiveness factor for the first time,
the analysis of the methods is not extensive.
A comprehensive study on popular methods in
PPDM accomplished in (Amiri and Quirchmayr,
2017). They defined a broader classification in both
categories (soft and traditional ways of privacy-
preserving) to reach an innovative idea about research
gaps and the state of the art of the topic. K-
Anonymity, Perturbation, Cryptographic &
Distributed methods, Association rule-based PPDM,
Classification based PPDM, Fuzzy Logic, Neural
Networks, Genetic Algorithms and Rough sets are the
main categories.
Analysis of state of the art shows that, among
popular methods used for privacy-preserving, soft
computing methods may yield better results. In this
paper four best popular soft methods selected for
designing the algorithms, including Neural Networks,
A Machine Learning Approach for Privacy-preservation in E-business Applications
445

Fuzzy sets, Genetic Algorithms and Rough Sets. This
selection is based on their unique characteristics and
demonstrated the capabilities of these methods.
Soft clustering is one of the most widely used
approaches in data mining with real-life applications.
Mingoti and Lima compared SOM and other
clustering algorithms like fuzzy c-means, k-means
according to their clustering capabilities on different
data structures. Their results show that there are
situations in which SOM performs well, even though
it is not the best (Mingoti and Lima, 2006). (Roh et
al., 2003) apply SOM clustering to their e-business
application in which empirical results demonstrate
that the SOM-based recommender system applied in
e-business designs supplies higher-quality referrals
than other comparative models. Hence, we also use
SOM clustering to enhance the privacy attribute while
obtaining the dual secondary goal of better accuracy
and performance.
3 PROBLEM DEFINITION
The privacy-preservation problem in e-business is
still an open problem and needs more secure methods
to improve the protection level of sensitive data.
However, improving privacy may account for
sacrificing the accuracy of data mining operations and
the efficiency of the overall process. So, proposed
methods should consider all these factors.
3.1 Problem Statement
The problem of privacy-preserving in data mining has
three viewpoints and, requires a solution which keeps
a balance in all the aspect defined in Figure 1:
- Improving privacy
- Keeping efficiency
- No information loss
1) Type of Environment
The focus of this paper lies on distributed
environments, since not only privacy and
optimization issues in this type of environments
seem to be more crucial, but also fewer studies
have been done on this definition. Moreover, with
the explosion of data, and requiring better
performance, resorting to distributed computing
and finding gaps and challenges of new issues in
this environment seems to be critical (Saxena and
Pushkar, 2017).
2) Type of Methods for Closing Privacy Gaps
Verykios and his partners demonstrated that
privacy-preserving problem might be an NP-hard
problem (Verykios et al., 2004b).The particular
capability of soft computing methods is their
power of deriving knowledge, as well as
extracting patterns and trends from complex data,
which are otherwise hidden in many applications
(Malik et al., 2012).
On the other hand, the comparison between two
factors of performance in Figure 3 and privacy
level demonstrates that none of the classic
methods even soft methods are satisfactory.
So, we will define our solutions with hybrid soft
computing methods. We will identify and discuss
new secure mining computations that enable
preserving privacy.
3) Selected Data Mining Task to Improve Privacy
As different data mining task may have different
feature and behaviour, to limit the scope of the
problem, soft clustering is selected to implement
the proposed privacy-preserving algorithms base
on the gaps that are have already found.
Figure 3: A comparison between two factors of privacy and
performance, in all methods, to show this point that none of
the soft nor hard methods are completely satisfying the PAE
factors.
We will augment a SOM clustering from different
aspects according to the PAE factors. The reason of
appointing SOM is based on the results of evaluating
clustering capability of SOM state where it was shown that
the SOM network is superior to hierarchical clustering
algorithms in e-business applications.
In this way, different proposed methods improve SOM
using a selection of different soft computing methods.
Some of our proposed methods are entirely new and make
use of procedures that have not been utilized in this context
before. Some others build on already existing techniques
Performance Privacy level
low
medium
high
SECRYPT 2018 - International Conference on Security and Cryptography
446

and improve them. In section 6, we explain the details of
these methods.
3.2 Methodology
The “design science research methodology” is used in
this paper which is typically applied to categories of
artifacts including algorithms, human/computer
interfaces, design methodologies (including process
models) and languages.
It is fundamentally a problem-solving paradigm.
Design science, as the other side of the IS research
cycle, creates and evaluates IT artifacts intended to
solve identified organizational problems. A
mathematical basis for design allows many types of
quantitative evaluations of an IT artifact, including
optimization proofs, analytical simulation, and
quantitative comparisons with alternative
designs(Von Alan et al., 2004).
We follow this methodology with these steps:
1) Identification of the criteria for the solution to
meet
2) Model development
3) Implementation
4) Test
5) Evaluation of results
6) Comparison with existing literature definitions
The first two parts discussed in this paper and
following parts will be complete in the future.
4 A PRIVACY -PRESERVING
SOM BASED MODEL - SPE
In this part, the fundamental concepts are discussed
which mainly work based on the Neural Networks as
a popular soft computing method and then, the SPE
model designed for this problem explained.
Figure 4: SOM network layers (Kohonen and Maps, 1995).
4.1 Prerequisite Definitions
Neural networks applied to solve the problems where
the solutions are either too complicated or impossible
to find a solution for a problem like privacy-
preserving that is uncertain and imprecise (Zadeh,
1994). Popular Neural Network models for data
mining purposes that seem suitable for the solution of
the mentioned problem include Hopfield networks,
Multilayer feedforward networks, and Kohonen’s
map.
The soft clustering method selected in this paper
to solve the privacy gaps is Kohonen’s map known as
SOM. We improve the PAE factors using soft
computing methods like fuzzy and GAs.
Kohonen’s self-organizing maps (SOM) is one of
the essential Neural Networks models for data mining
tasks, especially for data clustering and dimension
reduction (Kohonen, 1982). SOM can learn from
complex, multidimensional data and transforms them
into a topological map of much fewer dimensions,
typically one or two (Zhang, 2009). These
characteristics make SOM a suitable scheme to apply
for e-business applications, especially in
recommender systems.
A typical SOM network has two layers of nodes,
an input layer and an output layer called Kohonen's
layer. Each node in the input layer fully is connected
to nodes in the two-dimensional output layer. Figure
4 shows a simplified SOM network with several input
nodes in the input layer and a two-dimensional output
layer. SOM executes a series of iterations that each
consists of two phases: the competition and
cooperation phases (Kohonen and Maps, 1995). At
iteration t, during the competition phase, for each
input data X(t) = [Xi(t), X2(t), …, Xd(t)] from the
data set, the Euclidean distance between X(t) and
each neuron's
weight vector Wj(t) = [Wj,1(t),Wj,2(t), … , Wj,d(t)]
(1 ≤j ≤ K, where K is the total number of neurons in
the grid) is computed to determine the neuron closest
to X(t) as follows:




 






 




 







(1)
The neuron c with weight vector Wc(t) that has the
minimum distance to the input data X(t) is called the
winner neuron:






 

 (2)
A Machine Learning Approach for Privacy-preservation in E-business Applications
447

Figure 5: First plot of SPE model as a privacy-preserving SOM clustering.
During the cooperation phase, the weight vectors of
the winner neuron and the neurons in the
neighborhood G(rc) of the winner neuron in the SOM
grid are shared towards the input data, where rj is the
physical position in the grid of the neuron j. The
magnitude of the change declines with time and is
smaller for neurons that are physically far away from
the winner neuron. The function for change at
iteration t can be defined as Z(rj , rc, t)
The update expression for the winner neuron c and
the neurons in neighborhood G(rc) of the winner
neuron is shown as follows:








 















 

 (3)
SOM suffers from some privacy gaps in the
communication parts:
- securely discover the winner neuron from data
privately held by two parties
- safely update weight vectors of neurons
- firmly determine the termination status of SOM
(Han and Ng, 2007)
To overcome these challenge, some add-ons to
SOM designed to identify and address the privacy
gaps without information loss or significant overhead.
These hybrid techniques are based on soft methods
and will help to fix the breaches. The general
model(SPE) is shown in Figure 5. The yellow circles
in different parts of the picture indicate the proposed
methods which try to solve the issue. These methods
declared in following parts which try to improve
privacy with close a privacy breach that SOM suffer
from.
4.2 SPE Model
Privacy-preservation with attention to accuracy and
efficiency is an NP-hard problem (Bonizzoni et al.,
2009, Blocki and Williams, 2010). So, finding a
unique approach which could satisfy all these factors
is impossible. In this way, the SPE model proposed to
improve the privacy level of sensitive data regarding
other angles too. SPE consists of some independent
methods, each one tries to augment PAE factors.
These methods proposed to implement in different
parts of SOM transactions proposed according to the
privacy bugs founds. In this manner, each method
could focus to fix a privacy breach without any
distortion in common operations. In SPE some
independent methods are defined. Each method tries
to protect sensitive data without jeopardizing in the
system.
Method 1: Protecting sensitive data with
fuzzifying before training.
The idea of using fuzzy sets in the privacy-
preservation problem is trying to apply fuzzy
functions to preserve the private information of
individuals while revealing the details of the
aggregation in public. Results of analyzing related
works proved that fuzzy sets work for privacy
preservation and only introduce low additional costs
in the different cases that were explored above. Due
to its low cost, it may be applied in both types of
environments, especially in distributed servers.
However, using fuzzy sets as a single method of
privacy-preservation is not sufficient to cover new
attack(Li et al., 2017). Due to the progression of
attacks in prediction of the results of applying fuzzy
functions, using Fuzzy transforming function cannot
protect the sensitive data strongly. So, with using
fuzzy sets alone, there is no guarantee to reach the
best case of privacy-preservation. However, as a
SECRYPT 2018 - International Conference on Security and Cryptography
448

complementary approach, fuzzy methods are robust
tools.
Method 1. Fuzzify data before training
INPUT:
Dataset S consists of sensitive attribute data in m rows and
n columns.
OUTPUT:
Distorted Datasets that each S′ consist of m rows and n columns.
//Fuzzy membership functions such as Triangular-shaped,
S-shaped, and Gaussian used for transformation.
Begin
{
1.Suppress the identifier attributes
2.For each membership function (Triangular, Z-shaped, S-
shaped, Gaussian)
For each sensitive element in S do
{
Convert the element using selected
fuzzy membership function.
}
3. Release the all distorted datasets for training by SOM
function
}
End
Method 1 transforms the data set which consists of
some sensitive attributes. As discussed before, this
method is not enough, but in combining with other
methods in different parts of the model sounds useful
to implement. However, the idea of using fuzzy set in
this method returns to one of the SOM breaches that
mentioned in the previous section. This breach is the
high number of communication in the network. Fuzzy
function in this method transform the sensitive data
before training and prevent the bad guys to reach the
sensitive data. Protecting in this point of the model
(before training) is crucial at this stage data is raw and
with a high rate of probability bad guy can take
advantage of reaching them because of the high
number of distributed interactions. It makes a lot of
sense if in this point of model raw data transform to a
shape in which bad guy cannot recognize them
anymore.
Here, the sensitive data that defined by the user is
the input of method and, during an iteration function
try to fuzzify each item using a membership function
like Triangular, Z-shaped, S-shaped, Gaussian.
The output of the method would be a distorted
dataset that is ready for training by SOM function. In
this way, the worry of the high number of interaction
in the network reduced and, this because of applying
a method that distorted the data. As stated before, this
method for protection against privacy attacks is not
sufficient, but by applying other methods like method
3 and the idea of method 2, protection will be even
stronger.
Method 2: Implementing Offline SOM clustering.
The high amount of communication in SOM is a
considerable risk factor which exposes it to a variety
of security and privacy attacks. Decreasing the
amount of communication may diminish these risks.
In our proposed method which is based on
(Gorgonio and Costa, 2008), we apply extra add-ons
in order to decrease the risk of privacy attacks, and
also as a way to try to keep the performance of the
processes in the system.
The general plan is shown in figure 5 including
different offline SOM parties which train separately
and are then integrated with least steps.
Method 2. Offline SOM clustering
INPUT :
Raw Data to be trained and mapped including sensitive
information
OUTPUT :
Clusters consist of protected information
// Network Administrator decide about applying M1,M2,... based
on the level protection that is required
Begin
}
1. Repeat
}
each offline SOM train and map the data independently
// Arbitrary applying Method 1
add a new record to index vector for this offline SOM
{
Until the termination criterion is not satisfied .
// defined by Admin with number of offline SOM servers
2. integrating trained and mapped data into central SOM
3. arbitrary applying Method 2
4. training and mapping by central SOM
{
End
This method defines the overall process of the
SPE model. Its aim is reducing the number of
communications in a distributed network. So,
administrator divides the data in different local point
based on the e-business strategy. Each point acts
independently and, run the protecting methods like
method 1 to enrich itself against privacy breach. Next,
it runs SOM clustering and gets the clusters as the
results of the process. In this step, the local
administrator could decide about applying other
protecting methods like method 3 to be even more
protected. In this way, each local point has a unique
result which is the local clustering. Finally, the main
administrator asks local points to integrate trained and
mapped data into a central SOM. This integration by
itself runs on a secure channel and protected against
possible attacks.
A Machine Learning Approach for Privacy-preservation in E-business Applications
449

Method 3: Using Data Hiding with GA to Improve
Privacy Protection.
In a privacy-preserving concept, GAs (Genetic
Algorithms) can be applied as a complementary tool
in optimizing the results of primary algorithms. Most
of the approaches maximize the results of other
conventional methods of PPDM. According to our
research, when optimization is compulsory, soft
computing techniques, especially GAs, could yield
better results.
Moreover, in GAs, a chromosome resembles a
possible, feasible solution. Since the goal here is to
hide at most m appropriate transactions from a
database such that the fitness value can be optimal, a
chromosome with m genes is used, with each gene
representing a possible operation to hide.
Consideration of this method in our desired hybrid
model guarantees the accomplishment of all PAE
factors.
Method 3. Hiding sensitive data using GAs
INPUT:
A dataset S, consist of a set of sensitive itemsets, and the maximum
number m of transactions to be hidden, and a population size n.
OUTPUT:
An appropriate set of transactions to be hidden S′
//suitable fitness function to apply: Confidence based, Support
based and hybrid function
Begin
{
1. Derive the lower support threshold and Scan the database to
find the sensitive itemsets
2. Randomly generate a population of n individuals, with each gene
being the ID number of the transaction to be hidden
3. Repeat
{
1.Calculate the fitness value of each chromosome Ci in the
population // according to the selected fitness function
2. Execute the crossover and the mutation operations on the
population
3. Probabilistically choose individuals for the next generation
based on the selection scheme
}
Until the termination criterion is not satisfied.
4. create the output hidden transaction numbers in the best
chromosome to users
}
End
In this method, sensitive data defined by the user
is the input of methods. Firstly, a lower support
threshold set constantly and then the sensitive
itemsets derive from the dataset. This threshold
indicates the limit of deleting sensitive items set in
following steps. To start the operation of a GA
method usually a random population generate, but in
this method, this random population set with the ID
number of the sensitive transactions that defined in
the previous step.
The most important part of the method is the
fitness function that directly influences the usefulness
of the method. We are going to define this fitness
using some evaluation formulas used for checking a
privacy protection method. In the loop of GA method,
at step 2 the crossover and mutation operation execute
on the population. Population means the extracted
sensitive itemset. The last step of the loop chooses
the next generation based on a selection scheme. This
loop executes as long as the termination criteria met
which may be reaching an optimal set of sensitive
items or some extra factors.
5 DISCUSSION
Privacy-preserving in e-business comprises the issue
of big data and still is an open problem. Using a new
methodology like soft computing may bring much
needed improvements.
Neural Networks like SOM clustering is a
relatively newly suggested approach to privacy-
preservation and still needs to be worked on for two
central factors, optimizing the time for training the
network and privacy-preserving of its different parts.
Hybrid methods based on a Neural Network in
combination with other soft computing methods like
fuzzifying before training the neurons to close the
privacy gaps may yield better results.
The designed methods in the proposed model in
this paper have such a structure. The SPE model
described in Figure 5 provides the practitioner with
the ability to keep the balance between privacy
protection and performance. As discussed for Method
2, administrators decide about applying privacy
methods defined in different parts of the system. In
this way, whenever they want to improve the level of
protection, they should just enable the methods in
those parts of the system that may be confronted with
possible attacks.
First evaluations of simple and small datasets
show positive results on both sides, privacy
protection as well as performance. However, to
precisely evaluate the usefulness of this model,
further simulations with huge datasets are required to
prove the practical feasibility of the concepts. Thus,
in future work based on this model, independent
simulations using real data sets like the UCI machine
learning and the Movielens datasets will be used.
Moreover, extra soft computing methods such as
Deep Learning Neural Networks and Back
Propagation will be implemented, because of the
similarity in structure with the SPE models and they
will be compared with the currently proposed model.
SECRYPT 2018 - International Conference on Security and Cryptography
450

In case of positive results of the simulations, a further
comprehensive case study will implement the model.
Recommender systems that currently are a topic of
discussion in e-business applications from both sides,
privacy and utility, seem to be the most suitable case
study for evaluating the defined model.
All in all, the expected results of applying this
model, including various methods, would be sensitive
data better protected against privacy attacks in
different parts of SOM clustering when run in a
distributed environment. It should be noted that this
improvement of privacy protection would not
sacrifice the efficiency of regular SOM clustering.
Also, the amount of information loss would not be
considerable. In other words, the result of SOM
clustering before and after applying the proposed
methods in SPE model should be relatively similar in
terms of performance and accuracy.
6 CONCLUSION
The idea presented in this paper focuses on the issue
of privacy of individuals in e-business applications,
which involves big data and therefore data mining
techniques. Data mining on big data in combination
with privacy-preservation is still an open problem.
Among lots of methods proposed to improve privacy,
the lack of a strong method that could protect privacy
and also keeps efficiency and accuracy of the data
mining tasks at hand, still exists. Therefore, newer
methodologies like soft computing, also known as
machine learning, seem to be more useful for closing
these gaps. The proposed model in this paper
contributes to solving the defined problem in e-
business environments. The SPE model is flexible
and helps system administrators to keep a balance
between performance and privacy protection. Two
privacy-preserving methods have been introduced for
the SPE model which are independent and arbitrary
to implement. First results prove the usefulness of the
model and the methods, respectively. However more
simulations with huge datasets are still required to
check the utility of the SPE model in general. The
result of the proposed model in this paper is sensitive
data being protected against privacy attacks in SOM
clustering without significantly jeopardizing the
efficiency and accuracy of the general process.
ACKNOWLEDGEMENTS
This paper is partially funded by SBA Research,
Vienna, Austria.
REFERENCES
Aggarwal, C. C. & Philip, S. Y. 2008. A general survey of
privacy-preserving data mining models and algorithms.
Privacy-preserving data mining. Springer.
Agrawal, R. & Srikant, R. Privacy-preserving data mining.
ACM Sigmod Record, 2000. ACM, 439-450.
Amiri, F. & Quirchmayr, G. 2017. A comparative study on
innovative approaches for privacy-preserving in
knowledge discovery. ICIME ,ACM.
Blocki, J. & Williams, R. Resolving The Complexity of
some data privacy problems. International Colloquium
on Automata, Languages, and Programming, 2010.
Springer, 393-404.
Bonizzoni, P., Della Vedova, G. & Dondi, R. The k-
anonymity problem is hard. International Symposium
on Fundamentals of Computation Theory, 2009.
Springer, 26-37.
Chen, H. 2006. Intelligence and security informatics:
information systems perspective. Decision Support
Systems, 41, 555-559.
Fayyad, U. M., Piatetsky-Shapiro, G., Smyth, P. &
Uthurusamy, R. 1996. Advances in knowledge
discovery and data mining, AAAI press Menlo Park.
Gorgonio, F. L. & Costa, J. A. F. Combining parallel self-
organizing maps and k-means to cluster distributed
data. Computational Science and Engineering
Workshops, 2008. CSEWORKSHOPS'08. 11th IEEE
International Conference on, 2008. IEEE, 53-58.
Han, S. & Ng, W. K. Privacy-preserving self-organizing
map. DaWaK, 2007. Springer, 428-437.
Jahan, T., Narasimha, G. & Rao, C. G. 2014. A
Comparative Study of Data Perturbation Using Fuzzy
Logic to Preserve Privacy. Networks and Communica-
tions (NetCom2013). Springer.
Kohonen, T. 1982. Self-organized formation of
topologically correct feature maps. Biological
cybernetics, 43, 59-69.
Kohonen, T. & Maps, S.-O. 1995. Springer series in
information sciences. Self-organizing maps, 30.
Li, P., Chen, Z., Yang, L. T., Zhao, L. & Zhang, Q. 2017.
A privacy-preserving high-order neuro-fuzzy c-means
algorithm with cloud computing. Neurocomputing,
256, 82-89.
Maimon, O. & Rokach, L. 2007. Soft Computing for
knowledge discovery and data mining, Springer
Science & Business Media.
Malik, M. B., Ghazi, M. A. & Ali, R. Privacy preserving
data mining techniques: current scenario and future
prospects. Computer and Communication Technology
(ICCCT), 2012 Third International Conference on,
2012. IEEE, 26-32.
Mingoti, S. A. & Lima, J. O. 2006. Comparing SOM neural
network with Fuzzy c-means, K-means and traditional
hierarchical clustering algorithms. European Journal of
Operational Research, 174, 1742-1759.
Mitra, S., Pal, S. K. & Mitra, P. 2002. Data Mining in soft
computing framework: a survey. IEEE transactions on
neural networks, 13, 3-14.
A Machine Learning Approach for Privacy-preservation in E-business Applications
451

Narwaria, M. & Arya, S. Privacy Preserving Data Mining—
‘A state of the art’. Computing for Sustainable Global
Development (INDIACom), 2016 3rd International
Conference on, 2016. IEEE, 2108-2112.
Roh, T. H., Oh, K. J. & Han, I. 2003. The Collaborative
filtering recommendation based on SOM cluster-
indexing CBR. Expert systems with applications, 25,
413-423.
Samet, S. & Miri, A. 2012. Privacy-Preserving Back-
propagation and extreme learning machine algorithms.
Data & Knowledge Engineering, 79, 40-61.
Saxena, V. & Pushkar, S. Fuzzy-Based Privacy Preserving
Approach in Centralized Database Environment.
Advances in Computational Intelligence: Proceedings
of International Conference on Computational
Intelligence 2015, 2017. Springer, 299-307.
Verykios, V. S., Bertino, E., Fovino, I. N., Provenza, L. P.,
Saygin, Y. & Theodoridis, Y. 2004a. State-of-the-art in
privacy preserving data mining. ACM Sigmod Record,
33, 50-57.
Verykios, V. S., Elmagarmid, A. K., Bertino, E., Saygin, Y.
& Dasseni, E. 2004b. Association rule hiding. IEEE
Transactions on knowledge and data engineering, 16,
434-447.
Von Alan, R. H., March, S. T., Park, J. & Ram, S. 2004.
Design science in information systems research. MIS
quarterly, 28, 75-105.
Westin, A. 1999. Freebies and privacy: What net users
think.
Wu, X., Chu, C.-H., Wang, Y., Liu, F. & Yue, D. 2007.
Privacy preserving data mining research: current status
and key issues. Computational Science–ICCS 2007,
762-772.
Zadeh, L. A. 1994. Fuzzy logic, neural networks, and soft
computing. Communications of the ACM, 37, 77-85.
Zhang, G. P. 2009. Neural networks for data mining. Data
mining and knowledge discovery handbook. Springer.
SECRYPT 2018 - International Conference on Security and Cryptography
452