Privacy Preservation of Social Network Users Against Attribute
Inference Attacks via Malicious Data Mining
Khondker Jahid Reza
1
, Md Zahidul Islam
1
and Vladimir Estivill-Castro
2
1
School of Computing and Mathematics, Charles Sturt University,
Panorama Avenue, Bathurst 2795, NSW, Australia
2
Departament de Tecnologies de la Informaci
´
o i les Comunicacions, Universitat Pompeu Fabra,
Roc Boronat, 138, 08018 Barcelona, Spain
Keywords:
Attribute Inference, Data Mining, Privacy Protection Technique.
Abstract:
Online social networks (OSNs) are currently a popular platform for social interactions among people. Usually,
OSN users upload various contents including personal information on their profiles. The ability to infer users’
hidden information or information that has not been even uploaded (i.e. private/sensitive information) by an
unauthorised agent is commonly known as attribute inference problem. In this paper, we propose 3LP+, a
privacy-preserving technique, to protect users’ sensitive information leakage. We apply 3LP+ on a synthet-
ically generated OSN data set and demonstrate the superiority of 3LP+ over an existing privacy-preserving
technique.
1 INTRODUCTION
Data mining of users’ information on Online Social
Networks (OSNs) can reveal individuals’ private in-
formation. News pieces, reported by the Boston
Globe (Johnson, 2009), reveals that users’ sexual ori-
entation can be correctly predicted by accessing their
non-sensitive information. Hence, the ability of an
intruder to infer users’ sensitive information is known
as the attribute inference problem.
A technique (PrivNB), based on the Na
¨
ıve Bayes
classifier, can protect a sensitive attribute of users by
suppressing attribute values identified as predictors
and deleting some friendship links (Heatherly et al.,
2013). Another technique called 3LP (Reza et al.,
2017a) uses a decision forest algorithm to enhance
users’ privacy against the attribute inference prob-
lem. 3LP in its first layer suggests to a victim user
which predictor attribute values to suppress from the
user’s profile. If the sensitive information of the user
is still unprotected even after the suppression of pre-
dictor attributes, then 3LP suggests to hide some ex-
isting friends from the user’s friend list in its Layer 2
and add new friends in Layer 3.
In this study, we propose 3LP+ which is an ex-
tension of the existing 3LP algorithm (Reza et al.,
2017a). 3LP assumes the existence of a single sen-
sitive attribute such as the “Political View” while
in reality a user is likely to have multiple sen-
sitive attributes such as the “Political View” and
“Sexual Orientation”. To protect the privacy of mul-
tiple sensitive attributes 3LP could be applied mul-
tiple times, but every run of 3LP would be iso-
lated/independent. As a result, they can be counter-
productive in the sense that one run (say to protect the
“Political View”) might suggest hiding a friendship
information with another user while a subsequent run
(say to protect “Sexual Orientation”) might suggest
disclosing the same friendship information resulting
in the loss of protection of “Political View”.
3LP+ aims to provide privacy for multiple sensi-
tive attributes through a co-ordinated approach as op-
posed to the isolated approach. It uses a matrix to
store the history of any friendship being hidden or
new frinedship being created during a run to avoid
a conflicting suggestion in a subsequent run. For ex-
ample, if the t-th run suggests hiding a friendship of
the victim user with another user (and the victim user
actions on the suggestion), then the matrix stores that
information so that a subsequent run does not suggest
the victim user creating the friendship with the same
user.
The rest of this paper is organized as follows. In
Section 2, we describe the privacy attack model con-
sidered in this study. Section 3 presents the privacy-
preserving technique 3LP+. We describe our experi-
412
Reza, K., Islam, M. and Estivill-Castro, V.
Privacy Preservation of Social Network Users Against Attribute Inference Attacks via Malicious Data Mining.
DOI: 10.5220/0007390404120420
In Proceedings of the 5th International Conference on Information Systems Security and Privacy (ICISSP 2019), pages 412-420
ISBN: 978-989-758-359-9
Copyright
c
 2019 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

Table 1: A sample data set.
User Relationship status m relation Profession m profession m emotional Class Attribute
a Married 0 Businessman 0 0.56 Lonely (Lon)
b Widow 0 Businessman 0 0 Connected (Conn)
c Single 0.56 Labourer 0 0 Connected
d Married 0.51 Student 0.56 0.56 Lonely
e Single 0.51 Labourer 0 0 Connected
u Widow 0.51 Student 0.51 0 ?
(a) Tree-1 (b) Tree-2
Figure 1: Decision trees build on the sample data set given in Table 1.
mental set up in Section 4 and the experimental results
in Section 5. Section 6 gives a concluding remark.
2 PRIVACY ATTACK SCENARIO
An attacker M can follow any approach to invade
users’ privacy. We consider a data mining approach,
in this study, assuming that, M prepares a data set by
analyzing an OSN of N users, each with A attributes
and friendship information. We name an attribute, and
it’s corresponding value together as an attribute-value
pair. Here each attribute value (A
n
= v) is consid-
ered as a distinct binary attribute. In a Social attribute
network (SAN) model, both users and their attribute-
values are modeled as vertices.
The attacker can take advantage of a metric func-
tion as shown in Equation 1 (Adamic and Adar, 2003),
to incorporate the friendship information into D.
m(u, A
n
= v) =
∑
t∈Γ
s+
(u)∩Γ
s+
(A
n
=v)
w(t)
log|Γ
+
(t)|
. (1)
Here, Γ
s+
(u) is a set of OSN users connected to
a user u and Γ
s+
(A
n
= v) is the set of users hav-
ing the attribute-value A
n
= v. Similarly, Γ
A
n
+
(u) is
the set of all attribute-value pairs linked to user u.
Therefore, the neighbourhood of u is represented as,
Γ
+
(u) = Γ
s+
(u) ∪ Γ
A
n
+
(u). On the other hand, t is
the set of u’s friends who have an attribute-value pair
A
n
= v (i.e. t ∈ Γ
s+
(u) ∩ Γ
s+
(A
n
= v)) and Γ
+
(t) =
Γ
s+
(t) ∪ Γ
A
n
+
(t). The w(t) is the weight of each of
them and it’s value is set to 1 in this study. The higher
the value of m(u, A
n
= v), the higher the chance that u
has the value v for attribute A
n
.
For illustration, we suppose M wants to know
the emotional status of u who considers it as sen-
sitive and hence hides it from public view. To
launch the attack, M can first prepare a data set
by storing the user u’s available information (for
example, Relationship status and Profession) in it.
Then M can visit the profiles of other users who
disclose their emotional status as well as, additional
information (i.e. Relationship status and Profession
in this example) and store all these information
in D. We present a sample of such data set in Ta-
ble 1. The information directly related to OSN users
(e.g. Relationship status and Profession as shown
in Table 1) are named regular attributes, whereas
the information related to the friendship links (that
can be calculated by using Equation 1) are named
link attributes. M can consider “Emotional Status”
as the class attribute and apply any machine-learning
technique to obtain the patterns of the lonely and con-
nected users from the data set. In Figure 1, we present
a sample decision forest which can be built from D (as
shown in Table 1). Here, the rectangular boxes are
called nodes and ovals are called leaves. The path
from a root node to a leaf is called a logic rule. By us-
ing these rules, the attacker can predict the emotional
status of a new user u (who is not in Table 1) even
when the user does not disclose the information.
We argue that OSN users may have diverse pref-
erences on what they consider sensitive. For exam-
ple, one may consider their emotional status as sensi-
tive while others may consider their emotional status
Privacy Preservation of Social Network Users Against Attribute Inference Attacks via Malicious Data Mining
413

and/or political view as sensitive. Therefore, a
privacy-preserving technique should be capable of
protecting all sensitive information for each user and
should be capable of providing privacy when attack-
ers use a different classifier than the one that the tech-
nique expects them to use. Again, a large number
of attribute value suppression may provide better pri-
vacy, but at the same time, it also defeats the whole
purpose of using social network sites for a user. The
goal should be to provide privacy by suppressing the
minimum number of attribute values.
3 OUR TECHNIQUE
The basic idea of 3LP+ is to protect the privacy of all
information that a user consider sensitive. Users can
give the list of attributes they considers sensitive and
then, 3LP+ provides three steps (or layers) of recom-
mendations:
Step 1: Compute the sensitivity of each attribute
for each user and suggest to the user which attribute
values the user needs to suppress.
In Step 1, as shown in Algorithm , 3LP+ selects a
class attribute (from the list of sensitive attributes con-
sidered by the 3LP+ User u) randomly, prepares a
training data set D, and then applies SysFor (Islam
and Giggins, 2011) on D to get a set of logic rules.
3LP+ then uses the support and confidence of each
rule to compute its sensitivity (or Rule Sensitivity)
value in breaching the privacy of a sensitive attribute.
Similar to previous studies (Reza et al., 2017a; Reza
et al., 2017b), the rules having Rule Sensitivity value
1.006 or above are considered Sensitive Rules in this
study. We utilize the function GetSensitiveRules() to
represent the processes of generating the sensitive rule
set R
u
for u.
After preparing the sensitive rule set R
u
, 3LP+
counts the number of appearance of each regular at-
tribute A
n
in R
u
and store A
n
in A
s
. Here, A
s
stores all
the regular attributes and the number of their appear-
ances in R
u
. One attribute can appear only once in a
sensitive rule R
u
j
but more than once in R
u
. The reg-
ular attribute A
n
with the highest number of appear-
ances in R
u
is suggested to u for suppression. The
decision is then up to u whether to suppress its value
or not. If u suppresses the value of A
n
, then A
n
is no
longer available in A
s
and all sensitive rules in R
u
that
have A
n
in their antecedent are no longer applicable
for u. Regardless whether u suppresses the attribute
or not, 3LP+ then identifies A
n
with the next highest
appearances and suggests u to suppress that. The pro-
cess continues until R
u
or A
s
becomes empty.
Step 2: Hide friendship links as necessary if they
are not fabricated previously.
After Layer 1, if any sensitive rule remains in R
u
such a rule only uses link attributes (i.e. the attribute
values can only be altered by using Eq. 1). There-
fore, if a link attribute appears as an antecedent of
a sensitive rule R
u
j
(i.e. R
u
j
∈ R
u
), where the value
of the link attribute needs to be greater than a con-
stant SplitPoint (as mentioned in R
u
j
), the 3LP+ ex-
plores to reduce its value < SplitPoint by hiding some
of u’s friendship links. By doing this 3LP+ makes the
rule unusable to predict the class value of u with cer-
tainty.
In Step 2, 3LP+ first identifies the link at-
tribute, A
n
, that appears most in the sensitive rule
set R
u
and compute A
n
’s value, denoted as V , us-
ing Eq. 1. If V is higher than the split point men-
tioned in R
u
j
, then 3LP+ suggests u to hide a friend-
ship link. While choosing a friendship link, 3LP+ se-
lects a friend, t
i
, of u who has the smallest degree and
has not previously appeared in the friendship matrix F
(here F is an 1 × N matrix which stores the Flag in-
formation for u). The 3LP+ recommends u to hide t
i
so that it can reduce V ’s value the most by hiding a
minimum number of friends. If u follows the recom-
mendation, 3LP+ puts a Flag up in the i
th
column of
the friendship matrix F and this ex-friend will not be
recommended for further hiding or adding. 3LP+ then
updates G
0
, F
0
, and recomputes V ’s value.
This process continues until the value V is lower
than the SplitPoint mentioned in R
u
j
. Once the V is
lower than the split point, then the process of hid-
ing friends stops and 3LP+ removes R
u
j
and other
rules (which have an antecedent with the value V )
from R
u
as they are no longer be applicable to de-
termine u’s class value. At the end of Step 3, if R
u
is
not empty then only 3LP+ moves to Step 3 i.e. Layer
3.
Step 3: Add friendship links as necessary if they
are not fabricated previously.
After Step 1 and Step 2, any sensitive rule remains
in R
u
, that contains link attribute only and tests for
a value V ≤ some SplitPoint in its antecedent. In this
case, 3LP+ suggests u to add new friends so that the V
becomes greater than the split point in R
u
j
and thus R
u
j
is no longer applicable to u. While adding any friend
on u’s friend list, a user t
i
is selected in such a way that
a Flag has not been up previously in the i
th
column
of F
0
matrix and having the smallest Γ
+
(t) value. If u
accepts the recommendation, 3LP+ then updates the
matrix F
0
, friendship graph G
0
, and V increases. This
adding process continues until the value V exceeds the
split point value. It is noted that, adding a new friend
on a profile is complicated and depends on the other
users to confirm the friendship on OSN. Hence, 3LP+
ICISSP 2019 - 5th International Conference on Information Systems Security and Privacy
414

Algorithm 1: The Steps of 3LP+.
Input : User u, data set D, friendship network G, total number of records N in D, set of non-class attributes A, set of regular attributes A
r
, set of link
attributes A
l
where A
r
, A
l
⊂ A, set of class attributes C.
Output : Recommendations for u.
Variables : R = set of sensitive rules, R
j
= the j
th
sensitive rule, A
n
= the n
th
attribute, and F = 1 × N matrix stores F lag information for u /*Initially all
values in F are set to False */.
Step 1: Compute Sensitivity of Each Attribute for a User and Suggest the User to Suppress Attribute Values as Necessary.
R
u
← GetSensitiveRulesForUser(D,A,C, u)
foreach R
u
j
∈ R
u
do
n = 0 /* The value of n is always reset to 0 before the initiation of W hile loop */
while n < |A| do
if A
n
∈ A
r
AND A
n
is in the antecedent of R
u
j
then
A
s
← A
s
∪ {A
n
} /* Add A
n
in an array A
s
*/
Counter
n
← Counter
n
+ 1 /* Counts the number of appearance of A
n
in A
s
*/
n = n + 1
while R
u
6= φ AND A
s
6= φ do
n ← maxarg(Counter, A) /* Returns the index of the attribute that appears the most */
SuggestSuppress(A
n
) /* Suggest u to suppress the value of A
n
*/
if A
n
is suppressed then
A
s
← A
s
\ {A
n
}
R
u
← R
u
\ {R
u
j
} /* Rules using A
n
are removed from R
u
*/
Step 2: Hide Friendship Links as Necessary if they are not fabricated previously.
G
0
= G and F
0
= F
n ← FindIndexMostSensitive(A
l
, R
u
) and V ← CalculateValue(A
n
, u) /* using Equation 1 */
while R
u
6= φ AND A
l
6= φ do
for j = 1 to |R
u
| do
if A
n
is in the antecedent of R
u
j
AND V ≥ SplitPoint(R
u
j
, A
n
) then
while V ≥ SplitPoint(R
u
j
, A
n
) AND MoreFriends(u, G
0
, F
0
, A
n
) do
i ← FriendWithLeastDegree(u, G
0
, F
0
, A
n
) /* i is the index of the Friend with least degree when F
0
i
∈ F
0
is False */
SuggestHide(t
i
) and G
0
← HideLink(G
0
, u, t
i
) /* t
i
is the i
th
friend */
F
0
← Flag(F
0
,t
i
) /* F
0
i
is turned to True */
V ← CalculateValue(A
n
, u)
R
u
← R
u
\ {R
u
j
} /* Rules using A
n
are removed from R
u
*/
j=j+1
A
l
← A
l
\ {A
n
}
n ← FindIndexMostSensitive(A
l
, R
u
) and V ← CalculateValue(A
n
, u)
Step 3: Add Friendship Links as Necessary if they are not fabricated previously.
n ← FindIndexMostSensitive(A
l
, R
u
) and V ← CalculateValue(A
n
, u) using Equation 1 */
while R
u
6= φ AND A
l
6= φ do
for j = 1 to |R
u
| do
if A
n
is in the antecedent of R
u
j
AND V ≤ SplitPoint(R
u
j
, A
n
) then
while V ≤ SplitPoint(R
u
j
, A
n
) AND MoreUsers(G
0
, F
0
, A
n
) do
t
i
← UserWithLeastDegree(G
0
, F
0
, A
n
) /* i is the index of the User with least degree when F
0
i
∈ F
0
is False */
SuggestAdd(t
i
) and G
0
← AddLink(G
0
, u, t
i
) /* t
i
is the i
th
user */
F
0
← Flag(F
0
,t
i
) /* F
0
i
is turned to True */
V ← CalculateValue(A
n
, u)
R
u
← R
u
\ {R
u
j
} /* Rules using A
n
are removed from R
u
*/
j=j+1
A
l
← A
l
\ {A
n
}
n ← FindIndexMostSensitive(A
l
, R
u
) and V ← CalculateValue(A
n
, u)
keeps these recommendations as a last resort. Our
experimental results also indicate this step is seldom
required.
4 EXPERIMENTS
4.1 Data Set
We implement the privacy techniques on a syntheti-
cally generated OSN data set (Nettleton, 2015) and
denote it as D. The data set contains 1000 records, 11
regular attributes, and 50,397 friendship links among
the users. In order to insert users’ link attribute val-
ues into the data set, we calculate metric values for
each regular attribute (using Equation (1)) and there-
fore, the total number of attributes becomes 22 (i.e.
11 regular attributes and 11 link attributes).
We prepare three versions of D for the experi-
mental purposes for three different class attributes
and they are: political orientation denoted as D
P
,
Privacy Preservation of Social Network Users Against Attribute Inference Attacks via Malicious Data Mining
415

Figure 2: Distribution of 100 test data set records in each
fold.
religious view denoted as D
R
, and sexual orientation
denoted as D
S
. When we consider a particular at-
tribute as a class attribute then rest of the attributes are
considered as non-class attributes. For example, in D
P
data set, both religious view and sexual orientation
are considered as non-class attributes.
We follow 10-fold cross validation method
through out our experiments. Therefore, in each
fold, a training data set, D
tr
, contains 900 records
and a testing data set, denoted as D
ts
, contains 100
records (i.e. 10% of the total records). In order
to justify the efficiency of our technique, we again
split these 100 test data records, in each fold, into
three different groups based on the different percent-
age of records. We present these three groups and
their records distribution in Figure 2.
The Group 1 consists of 60 users who consider
any single attribute (i.e. either “political orientation”
or “religious view” or “sexual orientation”) as sensi-
tive. On the other hand, we assume 30 users, in Group
2, have considered any two attributes (out of the three
attributes) as sensitive. Finally, Group 3 consists of 10
users who consider all the three attributes as sensitive.
While preparing a test data set e.g. D
ts,X
we select
the records who consider X as a sensitive information
and return all other records in training data set D
tr,X
.
For example, we keep 50 records in D
ts,P
(as shown
in D
ts,P
column ) who consider political orientation
as sensitive and return rest of the 950 records in the
training data set D
tr,P
. Here the different cell colours
indicate different records and the same colour repre-
sents the same records.
4.2 Experimental Set-up
We now describe the experimental set-up in three
phases for three different sensitive attributes. In
Phase I, we first protect the privacy for political view,
then for religious view in Phase II, and finally,
for sexual orientation in Phase III. We argue that
the 3LP+ can protect privacy of all the sensitive in-
formation (which are selected by its users) regardless
to the sequence of selection as a class attribute. There-
fore, we also conduct experiments in an opposite se-
quence order but for simplicity we only describe the
experimental set-up here for first sequence order.
Phase I. At first step, shown in Figure 3, we prepare
a training data set D
tr,P
, and a testing data set D
ts,P
from the main data set D by considering users’
political view as a class attribute. At Step 2, we ap-
ply the two privacy preserving techniques, i.e. 3LP+
and PrivNB, on the insecure test data sets. Here the
term ‘insecure’ means that the 3LP+ or PrivNB have
not been applied previously on the test data sets and
hence the users’ class value can be determined by an
attacker easily. The test data sets are then secured by
the techniques, as shown in Step 3, denoted as D
0
ts,P
and D
∗
ts,P
respectively. We calculate and compare the
number of insecure users exists in the insecure and
secure data sets. In order to provide privacy 3LP+
and PrivNB modifies the data sets by hiding informa-
tion/friends or adding friends. Therefore, we use two
different symbols
0
and ∗ throughout the experimen-
tal set-up to denote the modified data sets by 3LP+
and PrivNB techniques respectively.
A privacy provider may not determine the classi-
fier which is going to be used by an attacker and there-
fore, the privacy protecting technique should be able
to protect privacy against any machine learning tools.
In our experiments we explore and compare the per-
formance of 3LP+ and PrivNB for different classifiers
such as Na
¨
ıve Bayes classifier (NB), Support Vector
Machine (SV M), and Random Forest algorithm (RF).
In order to do that we first apply these machine learn-
ing algorithms on insecure data set D
ts,P
in Step 1
and find the number of number of insecure users in
the test data set. We name it as classifiers’ accuracy
which refers to the number of users whose class value
is identified by the classifiers. The larger the accuracy
value indicates lower the privacy.
We then apply the different classifiers on secure
test data sets D
0
ts,P
and D
∗
ts,P
in Step 3. By comparing
the classifiers’ accuracy results, in Step 1 and Step
3, we then determine which technique provides bet-
ter privacy on the test data sets. The results are pre-
sented in terms of number of insecure users, denoted
as t
s
0
and classifiers’ accuracy, denoted as t
c
1
0
. In Step
3, we also calculate data utility (in terms of number
of suppressed attribute values) in D
0
ts,P
and D
∗
ts,P
after
applying the two privacy preserving techniques and
denoted as t
u
3
. In Figs. 3 - 10, we use different colours
of arrows to indicate the procedure of two different
privacy preserving techniques.
ICISSP 2019 - 5th International Conference on Information Systems Security and Privacy
416

Figure 3: Phase I of the experiments.
Figure 4: Prediction of class value accuracy of two privacy
preserving techniques.
Figure 5: Prediction of users’ class value(correctly) using
the same classifier used by the privacy preserving tech-
niques.
Phase II. After Phase I, we select religious view as
a class attribute. We first prepare training and test-
ing data sets and denote them as D
0
tr,R
and D
0
ts,R
which
are prepared from D
0
. Similarly, D
∗
tr,R
and D
∗
ts,R
are
prepared from D
∗
. In Step 5, different classifiers are
applied on D
0
ts,R
and D
∗
ts,R
, denoted by C
4
and C
5
re-
spectively, to measure the classifiers’ accuracy. Then
we apply 3LP+ and PrivNB on test data sets in Step
7 (in order to secure the users’ privacy). After Step 7
we again return all the records to original data set and
thus it modified to D
2
0
and D
2∗
for 3LP+ and PrivNB
respectively. In Step 9, we again investigate the safety
of users (who consider Political view as sensitive)
in D
0
ts,P
and D
∗
ts,P
due to providing the privacy to users
who consider Religious View again by analysing the
number of insecure users. in D
2
0
ts,P
and D
2∗
ts,P
only.
Phase III. We select sexual orientation as a class
attribute in this phase and similar to previous two
phases, we first prepare training and testing data sets
i.e. D
tr,S
and D
ts,S
as shown in Figure 10. In Step
10, we apply different classifiers, denoted by C
10
and
C
11
, on the two test data sets D
2
0
ts,R
and D
2∗
ts,R
to find
the classification accuracy before and applying any
privacy techniques. We then apply 3LP+ on D
2
0
ts,R
and PrivNB on D
2∗
ts,R
in Step 11. We denote D
3
0
ts,R
and D
3∗
ts,R
to represent the secure test data sets and ap-
ply different classifiers, denoted by C
10
and C
11
, again
on them in Step 12. After securing the test data sets,
similar to Phase I and Phase II, we again analyse and
compare the number of insecure users in Step 15a and
Step 15b as shown in Figure 10.
Privacy Preservation of Social Network Users Against Attribute Inference Attacks via Malicious Data Mining
417

Figure 6: Phase II of the experiments.
Figure 7: Performance of Na
¨
ıve Bayes in order to breach
users’ privacy in the test data sets.
Figure 8: Performance of Support Vector Machine in order
to breach users’ privacy in the test data sets.
5 EXPERIMENTAL RESULTS
We present and compare the experimental results
of two privacy preserving techniques i.e. 3LP+
Figure 9: Performance of Random Forest Algorithm in or-
der to breach users’ privacy in the test data sets.
and PrivNB in this section. The results are presented
in terms of accuracy (in percentage) and the step num-
bers (as mentioned in the experimental set-up Sec-
tion 4.2). Step number information is presented in
the x-axis and accuracy information is presented in
the y-axis. Here the term accuracy indicates the num-
ber of users whose class value is identified by the at-
tacker. Higher accuracy indicates the greater chance
for the intruder to infer the class value of a user and
vice versa. In Figure 4 we present the number of in-
secure users whose class value can still be inferred by
applying the same classifier used by the privacy pro-
tection technique. We first provide privacy by the two
privacy techniques separately as described in the Sec-
tion 4.2. Here y-axis represents prediction probability
to classify a record (regardless to correctly or incor-
rectly classified) by an intruder after the protection
ICISSP 2019 - 5th International Conference on Information Systems Security and Privacy
418

Figure 10: Phase III of the experiments.
techniques are applied. We observe that the probabil-
ity percentage of records is much higher for PrivNB
compared to 3LP+ except at Step 1. This is because
the privacy preserving techniques are yet to imple-
ment at Step 1 as shown in Figure 3. On the other
hand, in Figure 5, the percentage of correctly classi-
fied records by PrivNB is approximately 70% more
than the 3LP+.
In order to explore the performance of two pri-
vacy protection techniques against three conventional
classifiers i.e. NB, SV M, and RF, we use the avail-
able classifier packages in WEKA. Figure 7 shows
the results of using NB as a classification technique.
We observe when we provide privacy to users in test
data sets by using 3LP+, the classification accuracy
of NB drops about 10 percent as shown at Step 1 and
Step 3. However, this accuracy drops is less than
10 percent in case of PrivNB. The similar trend is
observed throughout the experimental steps and our
technique clearly outperforms the existing PrivNB
technique. On the other hand, a similar classification
accuracy drops is observed for SV M and RF classi-
fiers as shown in Figure 8 and Figure 9 respectively.
We also measure the data utility in terms of sup-
pression for both techniques. In each test data set, (50
records*10 regular attribute values=500) 500 maxi-
mum attribute values are available before applying
any privacy techniques. In Figure 11 we show the
results for Step 3, Step 7, and Step 12 only as the
Figure 11: Comparison of attribute value suppression.
privacy techniques are applied in these steps. We
observe, in each step, the number of attribute value
suppression by the PrivNB technique is three times
higher than 3LP+.
6 CONCLUSION
We propose 3LP+ in this study to provide users’
privacy on social media. Previous privacy preserv-
ing techniques can protect users’ single sensitive at-
tribute (from being inferred) whereas 3LP+ can pro-
tect users’ multiple sensitive attributes in one run. Our
experimental results indicate that 3LP+ can provide
better privacy while maintaining higher utility than
an existing privacy preserving technique even if an
attacker uses a different set of classifiers.
Privacy Preservation of Social Network Users Against Attribute Inference Attacks via Malicious Data Mining
419

REFERENCES
Adamic, L. A. and Adar, E. (2003). Friends and neighbors
on the web. Social networks, 25(3):211–230.
Heatherly, R., Kantarcioglu, M., and Thuraisingham, B.
(2013). Preventing private information inference at-
tacks on social networks. IEEE Transactions on
Knowledge and Data Engineering, 25(8):1849–1862.
Islam, Z. and Giggins, H. (2011). Knowledge discovery
through sysfor: a systematically developed forest of
multiple decision trees. In Proceedings of the Ninth
Australasian Data Mining Conference-Volume 121,
pages 195–204. Australian Computer Society, Inc.
Johnson, C. (2009). Project gaydar. The Boston Globe, 20.
Nettleton, D. F. (2015). Generating synthetic online so-
cial network graph data and topologies. In 3rd Work-
shop on Graph-based Technologies and Applications
(Graph-TA), UPC, Barcelona, Spain.
Reza, K. J., Islam, M. Z., and Estivill-Castro, V. (2017a).
3lp: Three layers of protection for individual privacy
in facebook. In IFIP International Conference on ICT
Systems Security and Privacy Protection, pages 108–
123. Springer.
Reza, K. J., Islam, M. Z., and Estivill-Castro, V. (2017b).
Social media users’ privacy against malicious data
miners. In Intelligent Systems and Knowledge Engi-
neering (ISKE), 2017 12th International Conference
on, pages 1–8. IEEE.
ICISSP 2019 - 5th International Conference on Information Systems Security and Privacy
420