,((($&0,QWHUQDWLRQDO&RQIHUHQFHRQ$GYDQFHVLQ6RFLDO1HWZRUNV$QDO\VLVDQG0LQLQJ

Privacy Preservation in Social networks through

alpha – anonymization techniques
Saptarshi Chakraborty, B.K.Tripathy
School of Computing Science and Engineering, VIT University, Vellore, Tamil Nadu, India 632014
[email protected], [email protected]
+918870685867, +919489343018
II. ANONYMIZED GRAPH CONSTRUCTION
Abstract— We propose an (α, k) anonymity model based on
the eigenvector centrality value of the nodes present in the raw A. Anonymous degree sequence generation
graph and further extend it to propose (α, l) diversity model and Firstly we compute the EVCS for the raw graph and then
recursive (α, c, l) diversity model which can handle the protection
convert it to a particular anonymous degree sequence
of the sensitive attributes associated with a particular actor. For
anonymization purpose, we applied noise node addition depending upon the anonymization model. To apply the
technique to generate the anonymized graphs so that the concept of α-anonymization on the social network structure,
structural property of the raw graph is preserved. Our proposed we chose the highest occurring sensitive label as the sensitive
methods add noise nodes with very minimal social importance. label with respect to which the α-deassociation property for
We applied eigenvector centrality concept over traditional degree that particular equivalence group will be calculated. This
centrality concept to prevent mixing of highly influential nodes modification makes the alpha-anonymization model more
with less influential nodes in the equivalence groups flexible rather than selecting a global sensitive attribute value.
1) (α, k) anonymous degree sequence generation
Keywords— social networks, alpha-anonymization, k-
anonymity, l-diversity, recursive (c, l) diversity, noise nodes
An EVCS sequence P is (α, k) anonymous degree
sequence if it satisfies the following three conditions – 1) All
the elements belonging to an equivalence group G are having
I. KEY TERMINOLOGIES the same degree i.e. G [i].d = G [i+1].d = … = G [i+ n].d,
where i=0, 2) Every equivalence group G should follow the α-
Eigenvector centrality based sequence (EVCS) – The EVCS of
deassociation property, 3) Every equivalence group G should
the nodes (n[1], n[2], …, n[m]) of a graph can be defined as
the non-decreasing order representation of the nodes based on satisfy k-degree anonymity. A graph satisfies (α, k) anonymity
their eigenvector centrality values i.e. (n[1].evc ≥ n[2].evc ≥ if the EVCS of that graph is an (α, k) anonymous degree
………≥ n[m].evc). sequence. For every equivalence group G in the (α, k)
anonymous degree sequence, the target degree for all the
α-deassociation Property - We borrowed the concept of α- nodes belonging to that equivalence group is the degree of the
deassociation property from [4], which was originally node which has the highest eigenvector centrality value.
proposed to handle relational micro-data anonymization. We 2) (α, l) diverse sequence generation
modified the original definition so that it can be applied on An EVCS sequence P is (α, l) diverse sequence if it is an
social network graph. Let (E, s) be the set of vertices in the (α, k) anonymous degree sequence and also if every
equivalence group E which are associated with sensitive label equivalence group individually satisfies l-diversity. A graph
s where α is a user specified threshold ranging between 0 to1, satisfies (α, l) diversity if the EVCS of that graph is a (α, l)
then the equivalence group satisfies α-deassociation for a diverse sequence. Our proposed (α, l) diverse sequence
particular α, if the relative frequency of s in that equivalence generation algorithm constructs equivalence groups from
group is less than or equal to α, i.e. (|(E, s)| / E) ≤ α . A social EVCS and then set the target degree for all the nodes present
network graph satisfies α-deassociation for a particular α if all in the raw graph. The target degree for all the nodes belonging
the equivalence groups satisfy the α-deassociation property. to an equivalence group is the degree of the node which has
(α, k) anonymity - A social network graph is (α, k) anonymous the highest eigenvector centrality value.
if all the equivalence groups satisfy k-degree anonymity and 3) Recursive (α, c, l) diverse anonymization
α-deassociation property. For any equivalent group G, if m different sensitive labels
are present, then the G satisfies recursive (c, l) diversity for a
(α, l) diversity - A social network graph is (α, l) diverse if all particular c and l value if f1 < c (fl + fl+1 +…….+ fm), where fi
the equivalence groups satisfy (α, k) anonymity and l-diversity denotes the number of occurrence of i-th most frequent label.
for a particular value of k and l. An equivalence group is recursive (α, c, l) diverse if the
equivalence group satisfies recursive (c, l) diversity and α-
(α, c, l) diversity - A graph satisfies (α, c, l) diversity for a
deassociation property. A graph is recursive (α, c, l) diverse if
particular α, if all the equivalence groups satisfy recursive (c,
all the equivalence groups satisfy the recursive (α, c, l)
l) diversity and α-deassociation property.
diversity individually. To generate the recursive (α, c, l)
diverse sequence, we used the concept of safety grouping

$621$0 $XJXVW3DULV)UDQFH 

$&0,6%1
'2,KWWSG[GRLRUJ
 ,((($&0,QWHUQDWLRQDO&RQIHUHQFHRQ$GYDQFHVLQ6RFLDO1HWZRUNV$QDO\VLVDQG0LQLQJ

Fig. 1: APL of (α, k) anonymized ER graph Fig. 2: APL of (α, l) anonymized graph Fig. 3: APL of recursive (α, c, l) anonymized graph

proposed in [3]. An equivalence group G satisfies safety noise nodes in such an intelligent manner that the social
grouping, if it follows the following three conditions- 1) |G| ≥ importance of the noise nodes are always very low as
k, 2) f1 < c (fl + fl+1 +…….+ fm), 3) ((f1 +1) / (f1(m-l+1))) < c. compared to the social status of the nodes which are already
For every equivalence group in the (α, c, l) diverse degree present in the raw graph. Our proposed algorithms also do not
sequence, the target degree for all the nodes belonging to that mix high influential nodes with low influential nodes. As the
equivalence group is the average of the degrees of all the method proposed in [3] also uses the noise node addition
nodes present in that equivalence group. technique to generate the anonymized graph, we compared the
results obtained from our method with that proposed in [3].
B. Graph construction algorithms We denote the method proposed in [3] as the KDLD method.
After generating the different anonymous degree sequences, We tested our proposed algorithms for k=5, 10, 15, and 20.
we applied the noise node addition technique proposed in [3] For (α, 5) diversity and (α, 10) diversity, except for k=20, our
to construct the anonymized graph. approach produces noise nodes with lesser average
eigenvector centrality value as compared to the average
III. RESULT AND ANALYSIS eigenvector centrality value of the nodes present in the raw
graph. For (α, 3, 10) diversity, we find that for all values of k,
A. Synthetic Dataset our proposed method generates a recursive (c, l) diverse graph
where the average eigenvector centrality value of the noise
We generated the Erdos Renyi (ER) model graph which
nodes is less than the average eigenvector centrality of the
consists of 1000 vertices and 5000 edges. This dataset is used
nodes present in the raw graph. So, if preservation of the
for (α, k) anonymization only. We compared the results social status of the users in the anonymized graph is a
obtained from our approach with the other existing k- parameter to determine the quality and performance of
anonymity techniques such as k-match algorithm [5], anonymization, then our proposed approach performs much
generalization method [1], against 1 neighbor [2]. An analysis better than the existing method which uses degree centrality
of the graph in Fig. 1, shows that except our method the APL concept to generate the anonymized graph.
difference between the raw graph and anonymized graph for
all the methods is significant. As we increase the value of k, IV. CONCLUSION
the APL of the anonymized graph generated from those
methods starts deviating more from the raw graph APL The proposed anonymization model performs better than the
whereas our proposed method produces anonymized graph existing k-anonymity models in preserving the structural
which have almost equal APL of the raw graph. property of the graphs. Our proposed algorithms also ensure
that the noise nodes added for anonymization purpose attain
B. Real Datasets low social importance. Estimation of utility of the anonymized
We used the co-authorship network data compiled by M. data and also the effect of the addition of noise nodes on the
Newman which consists of 1589 nodes and 2742 edges i.e. utility of the anonymized data can be studied further.
among those 1589 scientists, 2742 distinct connections
existed. We considered the first letter of the name of the References
scientists as the distinct sensitive label for our experiment [1] M. Hay, G. Miklau, D. Jensen, D. Towsley, and P. Weis, Resisting
purpose. From Fig. 2, we can see that as the k value is Structural Re-Identification in Anonymized Social Networks, Proc. VLDB
Endowment, 2008, vol. 1, pp. 102-114.
increased, our proposed method tends to generate anonymized [2] K. Liu and E. Terzi, “Towards Identity Anonymization on Graphs,”
graph with almost same APL as that of the raw graph. Fig. 3 SIGMOD ’08: Proc. ACM SIGMOD International Conference Management
shows the performance of our proposed recursive (α, c, l) of Data, pp. 93-106, 2008.
diversity algorithm to generate anonymized graph. The [3] M. Yuan, L. Chen, Yu, P.S. and T. Yu, Protecting Sensitive Labels in
Social Network Data Anonymization, Knowledge and Data Engineering,
deviation from the APL of the raw graph becomes minimal as IEEE Transactions on , vol.25, no.3, pp.633,647, March 2013
we increase the k value. [4] R. C. Wong and J. Li and A. W. Fu and K. Wang, {α, k)-anonymity: an
Since we used the noise node addition technique to enhanced k-anonymity model for privacy preserving data publishing, Proc.
generate the anonymized graph, so, we also analyzed the ACM SIGKDD, 2006, pp. 754-759.
effect of noise nodes on the final anonymized graph. It is [5] L. Zou, L. Chen, and M.T. Ozsu, “K-Automorphism: A General
Framework for Privacy Preserving Network Publication,” Proc. VLDB
observed that our proposed algorithms preserve the social Endowment, vol. 2, no.1, pp. 946-957, 2009.
status of the nodes present in the raw graph and also adds the