Today, more than two hundred Online Social Networks (OSNs) exist where each OSN extends to offer distinct services to its users such as eased access to news or better business opportunities. To enjoy each distinct service, a user innocuously registers herself on multiple OSNs. For each OSN, she defines her identity with a different set of attributes, genre of content and friends to suit the purpose of using that OSN. Thus, the quality, quantity and veracity of the identity varies with the OSN. This results in dissimilar identities of the same user, scattered across Internet, with no explicit links directing to one another. These disparate unlinked identities worry various stakeholders. For instance, security practitioners find it difficult to verify attributes across unlinked identities; enterprises fail to create a holistic overview of their customers. Research that finds and links disconnected identities of a user across OSNs is termed as identity resolution. Accessibility to unique and private attributes of a user like ‘email’ makes the task trivial, however in absence of such attributes, identity resolution is challenging. In this dissertation, we make an effort to leverage intelligent cues and patterns extracted from partially overlapping list of public attributes of compared identities. These patterns emerge due to consistent user behavior like sharing same mobile number, content or profile picture across OSNs. Translating these patterns into features, we devise novel heuristic, unsupervised and supervised frameworks to search and link user identities across social networks. Proposed search methods use an exhaustive set of public attributes looking for consistent behavior patterns and fetch correct identity of the searched user in the candidate set for an additional 11% users. An improvement on the proposed search mechanisms further optimizes time and space complexity. Suggested linking method compares past attribute value sets and correctly connect identities of an additional 48% users, earlier missed by literature methods that compare only current values. Evaluations on popular OSNs like Twitter, Instagram and Facebook prove significance and generalizability of the linking method.

1. Automated Methods for Identity Resolution
across Online Social Networks
Paridhi Jain
April 25th, 2016
Prof. Ponnurangam Kumaraguru (Advisor)
Prof. Alan Mislove (Northeastern University)
Prof. Amitabha Bagchi (IIT-Delhi)
Dr. Sachin Lodha (TRDDC)

2. Online Social Network (OSN)
“a pla&orm to build social rela2ons among people who share similar interests,
ac2vi2es, backgrounds or real-life connec2ons.” [Boyd et al.]
3
209 acPve OSNs in 2015
cerc.iiitd.ac.in

3. Coverage of Social Networks
4
• Unique Service
• At least 200 million
users register on
OSNs
• A user is bounded to
maintain mulPple
accounts

4. 5
Single User on Multiple OSNs!
cerc.iiitd.ac.in
Can we
predict the
link?
Can we find and link disconnected iden00es of a single user?
=
Iden0ty Resolu0on

5. Why do Identity Resolution?
6cerc.iiitd.ac.in

6. Enterprises: (De-duplicating audience)
Tip:
Create verified enterprise profile,
Campaign pages, product pages
and invite users to like / follow the
pages.
7cerc.iiitd.ac.in
Return of investment?
Calculate Social Audience

7. De-duplicating audience
Social audience = 437,632 + 153,000 + 805,097 or less??
8cerc.iiitd.ac.in

8. Security Practitioners (Attribute Aggregation)
“The Twi`er account has no
real name a>ached to it. But
Buzzfeed contributor found her
Tumblr iden0ty and idenPfied
the account owner as Shashank
***, a hedge fund analyst and
campaign manager.
False Sandy Update
Source: h`p://ediPon.cnn.com/2012/10/31/tech/social-media/sandy-twi`er-hoax/
9cerc.iiitd.ac.in

9. Challenges
cerc.iiitd.ac.in 10
Professional Opinion
DaPng
Heterogeneous OSNs
Personal
Degree of Details
Quality and descripPve personal
And professional informaPon
Li`le personal informaPon
DescripPve opinions
A>ribute Evolu0on
Time
InformaPon evolved on one but
not on other
{jainpari, Bangalore}
RegistraPon with same
informaPon on both OSNs
{paridhij, New Delhi}

10. Thesis Statement
A user’s iden22es across online social networks can be
searched and linked using past and present values of the
iden*fiable and discrimina*ve public aDributes.
Comparison using “Past and present values” take advantage
of a`ribute evoluPon
“IdenPfiable and public a`ributes” address challenge of
heterogeneous OSNs
11cerc.iiitd.ac.in

11. Formulation
followers.
An individual is denoted by I and her identity on a social network SNA is denoted by IA. The task
of identity resolution can be formally defined as follows.
Problem Definition 1: Identity Resolution: Given an identity IA of user I on social network
SNA, find her identity IB on social network SNB using a search function S and a linking function
L.
IB = max
1jN
(L(IA, IBj)) where IBj 2 S(IA))
Observing the two functions involved, the process of identity resolution in online social networks
can be divided into two subprocesses – identity search and identity linking. Identity search lists a
set of candidate identities on SNB, which are similar to the given known identity IA in accordance
to the search function S and are suspected to belong to user I. Such a set of candidate identities
is represented as S(IA) and its size is denoted by N. The search function S inputs IA’s attribute
value, a defined similarity metric simS, and search space (SNB in this scenario) as arguments, and
selects all identities (IB1 · · · IBj · · · IBN ) from the search space for whom similarity simS between
the candidate’s attribute value and IA attribute values is greater than a threshold. The threshold
12
12
Search Func2on S:
• Input: an idenPty, a search space
• Output: candidate set
Linking Func2on L:
• Input: an idenPty, a candidate set
• Output: Best matching candidate
cerc.iiitd.ac.in
Iden0ty Search
Iden0ty Linking
@janemargetkitchen*
Mrs.*Marget*
Cookie*Specialist**'Implies'comparison'of'complete'iden--es'
Figure 2.3: Architecture of an identity resolution process.
2.2.1 Identity Search
Problem Definition 2: For a user I, given her identity IA on social network SNA and a search
function S, find a set of identities IBj on social network SNB such that simS(IA, IBj) ✓, on
defined similarity metric simS and empirically calculated threshold ✓.
{IB1, . . . , IBj, . . . , IBN } = S(IA) s.t. simS(IA, IBj) ✓
Each identity IBj in the set is termed as candidate identity and the set as candidate set. The size
13
• Network attributes refer to the connections of the
followers.
An individual is denoted by I and her identity on a social
of identity resolution can be formally defined as follows.
Problem Definition 1: Identity Resolution: Given an
SNA, find her identity IB on social network SNB using a
L.
IB = max
1jN
(L(IA, IBj)) where
Observing the two functions involved, the process of iden
can be divided into two subprocesses – identity search an
set of candidate identities on SNB, which are similar to th
to the search function S and are suspected to belong to u
is represented as S(IA) and its size is denoted by N. Th
value, a defined similarity metric simS, and search space (
selects all identities (IB1 · · · IBj · · · IBN ) from the search
the candidate’s attribute value and IA attribute values is

12. Generic Identity Resolution
13cerc.iiitd.ac.in
Extract
available &
discriminaPve
features
Candidate
IdenPPes
IDENTITY SEARCH IDENTITY LINKING
Pairwise
Comparisons

13. My Contributions
– Iden0ty Search: Novel methods for creaPng candidate set by exploiPng
public and discriminaPve a`ributes; increase idenPty resoluPon
accuracy by 13%
– Iden0ty Linking: Novel method for effecPve linking idenPPes by
leveraging a`ribute history; reducPon in miss rate by 48%
14cerc.iiitd.ac.in

14. Identity Search
cerc.iiitd.ac.in 15
Extract
available &
discriminaPve
features
Candidate
IdenPPes
IDENTITY SEARCH IDENTITY LINKING
Pairwise
Comparisons
Aim: To retrieve a candidate set containing the idenPty we search for.

15. Formulation
16cerc.iiitd.ac.in
2.2.1 Identity Search
Problem Definition 2: For a user I, given her identity IA on social network SNA and a search
function S, find a set of identities IBj on social network SNB such that simS(IA, IBj) ✓, on
defined similarity metric simS and empirically calculated threshold ✓.
{IB1, . . . , IBj, . . . , IBN } = S(IA) s.t. simS(IA, IB) ✓
Each identity IBj in the set is termed as candidate identity and the set as candidate set. The size
13Search Func2on S:
• Can be computed with parPal informaPon
• Can be computed with different genre of informaPon (text, image)
@janemargetkitchen*
Mrs.*Marget*
Cookie*Specialist**'Implies'comparison'of'complete'iden--es'
Figure 2.3: Architecture of an identity resolution process.
2.2.1 Identity Search
Problem Definition 2: For a user I, given her identity IA on social network SNA and a search
function S, find a set of identities IBj on social network SNB such that simS(IA, IBj) ✓, on
defined similarity metric simS and empirically calculated threshold ✓.
{IB1, . . . , IBj, . . . , IBN } = S(IA) s.t. simS(IA, IBj) ✓
Each identity IBj in the set is termed as candidate identity and the set as candidate set. The size
13
• Network attributes refer to the connections of the
followers.
An individual is denoted by I and her identity on a social
of identity resolution can be formally defined as follows.
Problem Definition 1: Identity Resolution: Given an
SNA, find her identity IB on social network SNB using a
L.
IB = max
1jN
(L(IA, IBj)) where
Observing the two functions involved, the process of iden
can be divided into two subprocesses – identity search an
set of candidate identities on SNB, which are similar to th
to the search function S and are suspected to belong to u
is represented as S(IA) and its size is denoted by N. Th
value, a defined similarity metric simS, and search space (
selects all identities (IB1 · · · IBj · · · IBN ) from the search
the candidate’s attribute value and IA attribute values is

16. State of the art
– Only profile a`ributes (private and public) for IdenPty Search [Motoyama et al., Malhotra et. al.,
Liu et al.]
– LimitaPons of Profile Search -
– RestricPve search, owing to non-availability of common a`ributes across networks.
[Gender on Facebook, but not on Twi`er]
– Search with Limited a`ributes → Large candidate set size → Intensive IdenPty
Linking computaPons
– Users may choose different profile a`ributes → Miss out correct idenPty in the
candidate set
– Li`le research on using content and network a`ributes to search for candidate idenPPes
[consistent user behavior and not profile]
– Extensive use of both private and public a`ributes. Need user authorizaPon for idenPty
search
17cerc.iiitd.ac.in

17. Heuris0c Search on
available a>ributes
– Addresses the gap of literature by using
content and network idenPty search.
– Similarity based rules to find candidate
idenPPes matching with given idenPty
– Aim to improve recall
– Real-Time search
Unsupervised search on
discrimina0ve a>ributes
– Real-Pme approaches are
computaPonally and Pme expensive
(Search in the complete social network)
– Pre-segment the social network
– Reduces Pme complexity from O(n2) to
O(n)
18
Proposed Methods
cerc.iiitd.ac.in

18. Heuristic Identity Search
19cerc.iiitd.ac.in
Profile
Content
Self-mention
Network
Syntactic
and Image
Search Linking
If self-identified /
returned by
more than one
search method
No
Yes
Candidate
Identities
name,
location,
username
mobile no,
post,
friends,
followers
Paridhi Jain, Ponnurangam Kumaraguru, and Anupam Joshi. 2013. @I seek ‘L.me’: Iden2fying Users across Mul2ple Online Social
Networks. In Proceedings of the 22nd InternaPonal Conference on World Wide Web, WWW ’13 Companion. ACM, New York, NY, USA,
1259- 1268. DOI=h`p://dx.doi.org/10.1145/2487788.2488160 [Honorable MenPon Award}

19. 20
Content Search
Algorithm 2 Heuristic Search Methods
procedure Content Search
IA known identity on SNA
S {IA.source, IA.posts}
if S[0] 2 {HootSuite, TwitterFeed, Facebook} then
posts S[1]
for each m in posts do
remove stop-words and non-ascii characters from m
limi to 75 characters
query SNB API with m and retrieve candidates with similar posts
Cxs candidates
for each c in Cxs do
if sim(c.post, m)  0 then
delete c from Cxs
add Cxs to Cx
return Cxs
cerc.iiitd.ac.in

20. Evaluation
21
Ground Truth Dataset: 543 users from FriendFeed and
SocialGraph
Selec0on Strategy: Random selecPon
Why: To avoid any bias in evaluaPon. The methods are
produced to be generalizable.
Accuracy =
correctly identified
Total users
Precision =
Prelevant ∩ Pretrieved
Pretrieved
Recall =
Prelevant ∩ Pretrieved
Prelevant
Figure 3.1: Architecture of the identity resolution framework using proposed heuristic search methods and
linking methods from literature.
Table 3.2: Evaluation of the identity resolution framework with contribution of each search algorithm in
the resolution accuracy. Search methods based on profile (url), content, self-mention and network attributes
improve resolution accuracy by 13.1%.
Search Algorithm Ucorrect Accuracy
Profile Search (P) 205 37.7%
Content Search (C) 3 0.5%
Self-mention Search (SM) 31 5.7%
Network Search (N) 1 0.2%
Identity Search (P+C+SM+N) 220 40.5%
P (without URL) 149 27.4%
P (with URL) + (C+SM) + N 149+71 27.4% +13.1%
with the traditional profile search used in the literature, assuming access to only public profile
attributes. Traditional profile search method finds candidate identities by search parameters –
Improvised profile, content and network search methods successfully
improved the accuracy and the recall by 13.1%.
cerc.iiitd.ac.in

21. Unsupervised Identity Search
22
v/s complete
search space
v/s available
a>ributes
NiyaP Chhaya, Dhwanit Agarwal, Nikaash Puri, Paridhi Jain, Deepak Pai, and Ponnurangam Kumaraguru. 2015. EnTwine: Feature Analysis and
Candidate Selec2on for Social User Iden2ty Aggrega2on. In Proceedings of the 2015 IEEE/ACM InternaPonal Conference on Advances in Social
Networks Analysis and Mining, ASONAM ’15.

22. Find discriminative features
23
Class Majority
Index (CMI)
Match
No-Match
RaPo:
Encroachment
Index (EI)
DiscriminaPve if:
• Low
Encroachment
Index
• Low Error Index
Username Jaro Distance
Username LCS Distance
Username Levenshtein Distance
Username Character Bi-gram
Jaccard Index
Username Character Bi-gram
Cosine similarity
Name Jaro Distance
Name LCS Distance
Name Character Bi-gram Jaccard
Index
Name Character Bi-gram Cosine
similarity
Sample Features
cerc.iiitd.ac.in
Match:
{paridhi, paridhij}
No-match:
{paridhij,parineeta.c}
Error Index
(type-I/II)
error

23. 24
Modified Canopy Clustering
decreases to O(n). The search algorithm is modified and a concept of ‘sibling’ clusters is intro-
duced. As non–overlapping clustering tend to miss out some probable candidates, extending this
constrained set with siblings results in higher accuracy. The algorithm is given as Algorithm 6.
Algorithm 6 Modification to the Canopies
procedure Mod-Canopies
U set of user-profiles on the network
T threshold
d(x, y) distance measure
for each user-profile x in U :
create canopy Cx such that
for each user-profile y in U,
insert y into Cx if d(x, y) < T;
Remove all user profiles y added in the previous step from U.
loop while U is not empty;
The algorithm is similar to canopy clustering and its time complexity is still O(n2) in the worst
45
ModificaPons:
• Earlier overlapping canopies
• Overlapping canopies may not
reflect similarity with given
user idenPty
We create:
• Non-overlapping canopies
Discrimina-ve''
Features'
Iden--es'
IDENTITY'SEARCH' IDENTITY'LINKING'
@darkma'er_*
*John*Marget*
St.*Anthony*School*
@holy.james**
James*Marget*
St.*Anthony*School*
*
@dark.ma'er*
John*M*
New*Delhi*
.*
.*
.*
@janemargetkitchen*
Mrs.*Marget*
Cookie*Specialist*
(John,*John)**
.*
.*
*'Implies'comparison'of'complete'iden--es'
@dark.ma(er*
John*M*
New*Delhi*
Figure 2.3: Architecture of an identity resolution process.
2.2.1 Identity Search
Problem Definition 2: For a user I, given her identity IA on social network SNA and a search
function S, find a set of identities IBj on social network SNB such that simS(IA, IBj) ✓, on
defined similarity metric simS and empirically calculated threshold ✓.
{IB1, . . . , IBj, . . . , IBN } = S(IA) s.t. simS(IA, IBj) ✓
Each identity IBj in the set is termed as candidate identity and the set as candidate set. The size
13
cerc.iiitd.ac.in
paridhij
pari.nidhi
paridhijain
ridhi_jain
paritosh_jain Parineeta.jain
parineeta_joshi
r_jain
Raghav_jain
riju_
Discrimina-ve''
Features'
Iden--es'
IDENTITY'SEARCH' IDENTITY'LIN
@darkma'er_*
*John*Marget*
St.*Anthony*School*
@holy.james**
James*Marget*
St.*Anthony*School*
*
@dark.ma'er*
John*M*
New*Delhi*
.*
.*
.*
@janemargetkitchen*
Mrs.*Marget*
Cookie*Specialist*
(John,*James
(John,*John
.*
.*
*'Implies'comparison'of'complete'iden--es'
Figure 2.3: Architecture of an identity resolution process.
2.2.1 Identity Search
Problem Definition 2: For a user I, given her identity IA on social networ
function S, find a set of identities IBj on social network SNB such that sim
defined similarity metric simS and empirically calculated threshold ✓.
{IB1, . . . , IBj, . . . , IBN } = S(IA) s.t. simS(IA, IBj) ✓
Each identity IBj in the set is termed as candidate identity and the set as can
13
paridhij
pari.nidhi
paridhi
ridhi_jain paritosh_jain
Parineeta.jain
parineeta_joshi
r_jain
Raghav_jain
riju_
Algorithmic Pme complexity reduces to O(n)

24. 25
Unsupervised search for a candidate set
their distance. We experimente with different values of threshold T to determine the most optimal
one. With a very small value, we cannot be able to expand our candidate set since we will not find
any sibling clusters whereas with a extremely value, the candidate set can be too large making the
algorithm computationally expensive. The empirical threshold T for our dataset is set to 12.
Algorithm 7 Unsupervised search method
1: procedure Modified-Search
2: U User profile we are looking for
3: C set of non overlapping clusters
4: T threshold
5: d(Cx, Cy) distance measure
6: for each cluster Cx in C:
7: compute the distance d(U, Cx)
8: select cluster Cm such that
d(U, Cm) is minimum of all distances
computed above, this is the most suitable cluster;
9: L List of suitable clusters, initially empty
10: for each cluster Cx in C:
11: if d(Cm,Cx) < T then
if d(U, Cx) < T then
append Cx to L
12: L holds our list of candidate clusters
For search, look for:
• Sibling Canopies
• Similar to most suitable
canopy AND similar to the
searched user profile
cerc.iiitd.ac.in

25. Evaluation
26
M = Match class; NM = No-Match Class
# of Users
(M:NM::1:1)
Threshold Precision
(Canopy)
Recall
(Canopy)
Precision
(MOD-
Canopy)
Recall
(MOD-
Canopy)
20000 0.95 0.15 0.90 0.25 0.79
20000 0.97 0.20 0.70 0.30 0.55
20000 0.98 0.24 0.62 0.33 0.69
Increasing the threshold, increases precision, degrades recall
Facebook-Twi`er

26. So far…
cerc.iiitd.ac.in 27
@darkmaDer_
J Marget
St. Anthony School
@holy.james
James Marget
St. Anthony School
@dark.maDer
John M
New Delhi
.
.
.
@janemargetkitchen
Mrs. Marget
Cookie Specialist
IdenPty
Search

27. Identity Linking
cerc.iiitd.ac.in 28
Extract
available &
discriminaPve
features
Candidate
IdenPPes
IDENTITY SEARCH IDENTITY LINKING
Pairwise
Comparisons
Aim: To retrieve best among the candidate set, i.e. the correct idenPty of the user

28. Formulation
29
little has contributed to address these challenges and drawbacks of profile search.
2.2.2 Identity Linking
Problem Definition 3: Given an identity IA of user I on social network SNA, a set of candidate
identities Q = S(IA) = {IB1, . . . , IBj, . . . , IBN } on social network SNB and a linking function L,
locate an identity pair (IA, IBj) such that L(IA , IBj) = max{L(IA, IB1),. . . , L(IA, IBN )}. IBj with
highest link-score is inferred as IB.
IB = max
1jN
(L(IA, IBj)) where IBj 2 Q)
An identity linking method estimates the correspondence between identity IA and each candidate
identity IBj by calculating a link-score L(IA, IBj) between their respective attributes and then rank
the candidate set on the basis of link-score. Candidate identity IBj with highest link-score is con-
cluded, as IB. The function L can be computed for all variety of data – text, date, image and
location. The function can either be a supervised classifier decision boundary or a heuristic rule, in
both scenarios, the function can be computed with partial and complete information.
cerc.iiitd.ac.in
Linking Func2on L:
• Can be a rule or a supervised classifier
• Can be computed with parPal informaPon
• Can be computed with different genre of informaPon (text, image)
New*Delhi*
.*
.*
.*
@janemargetkitchen*
Mrs.*Marget*
Cookie*Specialist**'Implies'comparison'of'complete'iden--es'
Figure 2.3: Architecture of an identity resolution process.
2.2.1 Identity Search
Problem Definition 2: For a user I, given her identity IA on social network SNA and a search
function S, find a set of identities IBj on social network SNB such that simS(IA, IBj) ✓, on
defined similarity metric simS and empirically calculated threshold ✓.
{IB1, . . . , IBj, . . . , IBN } = S(IA) s.t. simS(IA, IBj) ✓
Each identity IBj in the set is termed as candidate identity and the set as candidate set. The size
13
• Content attributes describe the content she creates o
post.
• Network attributes refer to the connections of the
followers.
An individual is denoted by I and her identity on a social
of identity resolution can be formally defined as follows.
Problem Definition 1: Identity Resolution: Given an
SNA, find her identity IB on social network SNB using a
L.
IB = max
1jN
(L(IA, IBj)) where
Observing the two functions involved, the process of iden
can be divided into two subprocesses – identity search an
set of candidate identities on SNB, which are similar to t
to the search function S and are suspected to belong to u
is represented as S(IA) and its size is denoted by N. Th
value, a defined similarity metric simS, and search space
selects all identities (IB1 · · · IBj · · · IBN ) from the search

29. State of the art
– Methods link idenPPes using
– Profile a`ributes [Zafarani et al., Perito et al., Malhotra et al., Liu et al. ]!
– Content a`ributes [Iofciu et al., Liu et al., Goga et al.]!
– Network a`ributes [Bartunov et al., Narayanan et al., Labitzke et al.]!
– Crowd sourced mechanisms [Shehab et al.]!
– Search Engines [Bilge et al.]!
– Most literature methods assume, compare and match access to
present (current) a`ributes of the idenPPes.
– But, current versions of the idenPPes may fail to match due to
– User choice
– A`ribute EvoluPon
30cerc.iiitd.ac.in

30. User choice
A private user may consciously choose to
de-link her idenPPes across OSNs, hence
current versions display different
personaliPes of the same user
A>ribute Evolu0on
An acPve user may keep on evolving their
a`ributes to suit trends, requirements, or
purpose. Thus, the current versions may
differ
31
Why current versions may fail to match
Username Name Descrip. Location Lang. Zone ProfilePic
0
10
20
30
40
50
60
70
%ofusers
2 values
3 values
4 values
5 values
cerc.iiitd.ac.in

31. Proposed Identity Linking
– If current versions do not match and if the user behavior is consistent across OSNs, any
of the past versions “may” match.
32
Supervised
Classification
Feature: 1
Feature: n
Similarity: 1
Similarity: n
Patterns of username creation behavior across OSNs
Patterns of username reuse behavior across OSNs
.
.
.
.
.
.
Labeled
datasets
US: {‘eenjolrass',‘isabelnevills',
‘giuliettacapuleti',‘tobsregbo'}
UC: {‘enjoolras',‘isabelnevilles'}
uc: {‘isabelnevilles'}
SNA
SNB
Feature: 1
Feature: m
Similarity: 1
Similarity: m
.
.
.
.
.
.
Username Sets PredicPon
3Paridhi Jain, Ponnurangam Kumaraguru, and Anupam Joshi. 2015. Other Times, Other Values: Leveraging ADribute History to Link User
Profiles across Online Social Networks. In Proceedings of the 26th ACM Conference on Hypertext & Social Media, HT ’15. ACM, New York,
NY, USA, 247-255. DOI=h`p://dx.doi.org/10.1145/2700171.2791040.

32. Username Set Collection
33
Tumblr username on the URL
Twi`er username
33
cerc.iiitd.ac.in
• Past usernames:
• Automated Tracking System that queries a user’s ID via API to
record her changed profile a`ributes
• her username on the OSN
• her URL a`ribute signifying change to her other OSN username
• Old Twi`er URL – abcd_efgh.tumblr.com
• New Twi`er URL – xyz.tumblr.com
• Ground Truth:
• Self-idenPficaPon behavior [Cross-referencing one’s OSN accounts]

33. Example
– User ID: 595**942*
– Past usernames on Twi>er:
– ["bigeasye_", "reezy11_", "epiceric_", "soulanola", "swampson_", "hebetheeeric",
"swampkidd_"]
– Past Usernames on Tumblr:
– ["bigeasye_", "epiceric17", "swampson", "hebetheeeric"]}
34cerc.iiitd.ac.in

34. Methodology
35
Supervised
Classification
Feature: 1
Feature: n
Similarity: 1
Similarity: n
Patterns of username creation behavior across OSNs
Patterns of username reuse behavior across OSNs
.
.
.
.
.
.
Labeled
datasets
US: {‘eenjolrass',‘isabelnevills',
‘giuliettacapuleti',‘tobsregbo'}
UC: {‘enjoolras',‘isabelnevilles'}
uc: {‘isabelnevilles'}
SNA
SNB
Feature: 1
Feature: m
Similarity: 1
Similarity: m
.
.
.
.
.
.
Collected Username Sets PredicPon
cerc.iiitd.ac.in

35. Features
Username Set Similarities
Syntactic
Static Creation
Similar Length
Similar Choice of Characters
Similar Arrangement of Characters
Evolutionary Creation
Stylistic
Occasional Reuse
Common username?
Best similarity score
Second Best similarity score
Frequent Reuse
Common username set
Temporal ordering?
Temporal sync?
Evolution of Length
Evolution of Choice of Characters
Evolution of Arrangement of Characters
Temporal
Case
LeetSpeak
Emphasizer
Prefix / Suffix
Slang words
Bad words
Function words
Phonetic Replacement
Grammar
36cerc.iiitd.ac.in

36. Evaluation
37
Supervised
Classification
Feature: 1
Feature: n
Similarity: 1
Similarity: n
Patterns of username creation behavior across OSNs
Patterns of username reuse behavior across OSNs
.
.
.
.
.
.
Labeled
datasets
US: {‘eenjolrass',‘isabelnevills',
‘giuliettacapuleti',‘tobsregbo'}
UC: {‘enjoolras',‘isabelnevilles'}
uc: {‘isabelnevilles'}
SNA
SNB
Feature: 1
Feature: m
Similarity: 1
Similarity: m
.
.
.
.
.
.
Collected Username Sets PredicPon
cerc.iiitd.ac.in

37. Datasets
– Linking profiles
– Twi`er – Tumblr
– Twi`er – Facebook
– Twi`er – Instagram
– Past usernames available for both profiles:
– 18,959 posiPve pairs, 18,959 negaPve pairs
– Past usernames available only on Twi`er but current username available on other
profile:
– 109,292 posiPve pairs, 109,292 negaPve pairs
38cerc.iiitd.ac.in
Network-Pair Twi>er-Tumblr Twi>er-Facebook Twi>er-Instagram Total Users
History on both 14,301 1,166 3,492 18,959
History on source only 58,285 31,076 19,931 109,292

38. 1. Independent Supervised Framework
2. Fusion Supervised Framework
Supervised Classification
39
3. Cascaded Supervised Framework
Classifier I
Current Username Features
[Exact Match, Substring Match]
Classifier II
Username Set Features
[Naive Bayes, SVM, DecisionTree, Random Forest]
Negative?
Positive?
Same User
Different Users
Negative?
US: {‘eenjolrass',‘isabelnevills',
‘giuliettacapuleti',‘tobsregbo'}
UC: {‘enjoolras',‘isabelnevilles'}
uc: {‘isabelnevilles'}
{‘tobsregbo'
‘isabelnevilles}
Us - UC
(or US - uc )
cerc.iiitd.ac.in

39. Prediction
40
Framework Config. Accuracy FNR FPR
Exact Match (b1) 55.38 89.34 0.00
Substring Match (b2) 60.99 78.46 0.00
Independent [b1→Naive Bayes] 72.10 53.81 1.91
Fusion [b1→Naive Bayes] 72.93 51.89 0.19
Cascaded [b1→Naive Bayes] 73.12 48.87 3.07
Cascaded [b1 → SVM [Linear]] 76.97 40.87 3.71
Cascaded [b2 → Naive Bayes] 73.27 48.52 3.14
Cascaded [b2 → SVM [Linear]] 76.93 40.87 3.78
- 48.47% cerc.iiitd.ac.in

40. So far…
cerc.iiitd.ac.in 41
@darkmaDer_ @holy.james
@magascus, @hello_kiDy
@darkmaDer_
@hello_kiDy, @magascu_,
@holy.james
Exis0ng iden0ty linking
Same user
Different users
Proposed iden0ty linking
Reduc0on of
FNR from
89% to 40%

41. cerc.iiitd.ac.in 42
Extract
available &
discriminaPve
features
Candidate
IdenPPes
IDENTITY SEARCH IDENTITY LINKING
Pairwise
Comparisons
Uses A`ribute EvoluPon Uses a`ributes that are shared
across OSNs
Proposed methods exploit…

42. A`ribute EvoluPon
– Implies: Out of sync idenPPes in Pme
– IdenPfy possible reasons and
characterisPcs
– ImplicaPons?
A`ribute Sharing
– Implies: sharing sensiPve informaPon
– IdenPfy possible reasons and
characterisPcs?
– Risks? Privacy ImplicaPons? Do users
care?
43
Understanding …
cerc.iiitd.ac.in

43. Attribute Evolution
– Aim: To understand how, why, and what fracPon of users have “out-of-sync”
idenPPes across OSNs
– Tracked about 8.7 million random Twi`er users and analyzed 10K users in
depth who evolved over Pme [selecPve sampled]
– Studied a unique idenPfiable public a`ribute - username
– Observa0ons:
– 20% of users consPtute 80% of username changes observed on Twi`er
– New usernames are disPnctly different from the old usernames
– A secPon of these users change for benign reasons like space gain, change of
idenPfiability while others are suspected with malicious intenPons
– Implica0on: Due to a`ribute evoluPon, quality dataset of past idenPPes of a
user is available. This instead of a challenge, becomes an opportunity for our
proposed idenPty linking.
cerc.iiitd.ac.in 44

44. Attribute Sharing
– Aim: To understand the reasons and risks of sharing sensiPve idenPfiable
informaPon about oneself
– Collected 2,492 Indian mobile numbers from OSNs like Twi`er and Facebook
public posts, bio and name
– Observa0ons:
– Mobiles numbers are pushed across mulPple OSNs, intenPonally and
unintenPonally
– Publicly shared sensiPve informaPon like mobile number can expose idenPfiable
details (ID, name, family) if collated with external data sources
– Implica0ons:
– Awareness of collaPon risks associated with sensiPve sharing is necessary.
Technological soluPons should a`end to it.
– Sharing sensiPve informaPon can implicitly resolve idenPPes
cerc.iiitd.ac.in 45

45. Contributions Summary
– Methods for idenPty search that exploit public a`ributes and user
behavior across OSNs
– We address the challenge of heterogeneous OSNs by considering only public and
universally available a`ributes
– Method for idenPty linking that leverage user evoluPon over Pme
– We exploit the challenge of a`ribute evoluPon to our advantage. Compare both
past and current versions of the idenPPes
– Observed and characterized user behavior that aids our proposed
methods
– We add to exisPng knowledge for development of our methods as well as future
idenPty resoluPon methods
46cerc.iiitd.ac.in

46. Implications to?
– Enterprises can carry out:
– Automated audience de-duplicaPon
– Automated psychographic segmentaPon based on aggregated user
profiles and inferred a`ributes.
– Security pracPPoners can de-anonymize malicious users
– Users
– Can be`er understand their idenPty leaks and patch them to avoid
idenPty resoluPon
– E.g. “should not share same content”, “should not create similar histories
of username”
– Risks of sharing sensiPve informaPon needs to the communicated by new
Over-the-top (OTT) applicaPons
47cerc.iiitd.ac.in

47. Limitations and Future Work
– Dependency on API
– LimiPng to only usernames for idenPty linking
– EvaluaPon on self-idenPfied users
– Future work:
– Extend to include past versions of idenPPes for be`er idenPty
search methods
– Extend to exploit evoluPon of mulPple a`ributes in a Pme
synchronized manner for idenPty linking
– Develop an OTT messenger that highlights possible leaks of
sensiPve informaPon, privacy and idenPty to a user
48cerc.iiitd.ac.in

48. Peer-reviewed Publications (1)
– Paridhi Jain, Ponnurangam Kumaraguru, and Anupam Joshi. 2013. @I seek ‘L.me’:
Iden2fying Users across Mul2ple Online Social Networks. In Proceedings of the 22nd
InternaPonal Conference on World Wide Web, WWW ’13 Companion. ACM, New York,
NY, USA, 1259- 1268. DOI=h`p://dx.doi.org/10.1145/2487788.2488160
– NiyaP Chhaya, Dhwanit Agarwal, Nikaash Puri, Paridhi Jain, Deepak Pai, and
Ponnurangam Kumaraguru. 2015. EnTwine: Feature Analysis and Candidate Selec2on for
Social User Iden2ty Aggrega2on. In Proceedings of the 2015 IEEE/ACM InternaPonal
Conference on Advances in Social Networks Analysis and Mining, ASONAM ’15. ACM,
New York, NY, USA, 1575-1576, DOI=h`p://dx.doi.org/10.1145/2808797.2809340.
– Paridhi Jain, Ponnurangam Kumaraguru, and Anupam Joshi. 2015. Other Times, Other
Values: Leveraging ADribute History to Link User Profiles across Online Social Networks.
In Proceedings of the 26th ACM Conference on Hypertext & Social Media, HT ’15. ACM,
New York, NY, USA, 247-255. DOI=h`p://dx.doi.org/10.1145/2700171.2791040.
49cerc.iiitd.ac.in

49. Peer-reviewed Publications (2)
– Paridhi Jain and Ponnurangam Kumaraguru. 2016. On the Dynamics of Username
Changing Behavior on TwiDer. In Proceedings of the 3rd IKDD Conference on Data
Science, 2016, CODS ’16. ACM, New York, NY, USA, ArPcle 6 , 6 pages. DOI=h`p://
dx.doi.org/10.1145/ 2888451.2888452.
– Prachi Jain, Paridhi Jain, and Ponnurangam Kumaraguru. 2013. Call me Maybe: Un-
derstanding Nature and Risks of sharing Mobile Numbers on Online Social Networks. In
Proceedings of the first ACM Conference on Online social networks, COSN ’13. ACM,
New York, NY, USA, 101-106, DOI=h`p://dx.doi.org/10.1145/2512938.2512959.
– Paridhi Jain, Tiago Rodrigues, Gabriel Magno, Ponnurangam Kumaraguru, and Virgilio
Almeida. Cross-Pollina2on of Informa2on in Online Social Media: A Case Study on Popular
Social Networks. In Proceedings of the 2011 IEEE 3rd InternaPonal Conference on Social
CompuPng, SocialCom ʹ11, pages 477–482, Oct 2011.
50cerc.iiitd.ac.in

50. Acknowledgments
51
• My advisor ‘PK’
• Prof. Anupam Joshi and Prof. Rahul Purandare
• Members of Precog@IIITD and CERC@IIITD
• Supported by TCS Research Fellowship (2010 – 2016)
• Friends, Colleagues and Family
Niharika Siddhartha AdiP Prateek Anupama SrishP cerc.iiitd.ac.in

51. Thanks!
Paridhi.jain@xerox.com
52cerc.iiitd.ac.in