The Friend-of-a-Friend Vocabulary (FOAF) is used in many online social networks to represent information about users and their friendships. Previous work has looked at how FOAF can be used to merge accounts across social networks. We often consider that merging as a bene t to the user, connecting their personal information and friend lists which would otherwise stay isolated. However, users may create multiple accounts - even on the same network - to intentionally separate their data. FOAF makes it equally easy to merge these accounts. In this paper, we are interested in the impact Semantic Web reasoning - with no additional data mining technology - can be used to resolve multiple accounts and the implications that has for privacy and safety online. We crawled FOAF pro les from all the social networking website that generate it, and looked at the pro les of individuals with multiple accounts to understand how they were using these accounts and why they were created. We present the results of this analysis and discuss the implications.
Just as social networks are one of the most popular web-based activities FOAF is one of the most widely used ontologies on the Semantic Web. With its semantics and the application of a Semantic Web reasoner, it is possible to merge pro les and friendship connections within and across social networking websites. While this capability is generally viewed as one of if not the greatest powers and bene ts of the Semantic Web, it also has implications for privacy and security. Users may want to create multiple pro les to keep parts of their lives separate, and this type of reasoning would impinge on the privacy they expect. At the same time, these multiple accounts could be used for nefarious purposes, and the reasoning and merging may allow us to detect and prevent this bad behavior.
In this paper, we ask one core question: Using Semantic Web reasoning on existing FOAF les, what can we learn about the frequency of, reasons for, and use of multiple pro les belonging to the same person in online social networks? While there are mechanisms for generating FOAF when it is not available and for merging pro les outside of Semantic Web reasoning, our work is a study of this problem on the Semantic Web as it exists today. Thus we are interested only in Semantic Web technologies and existing Semantic Web data. In addition to focusing our results on a speci c problem, this restriction means that the conclusions we draw are immediately applicable because they cover existing data analyzed with existing techniques. 2 2.1
The FOAF vocabulary1 is used to describe people, their attributes, and the relationships between people. FOAF is written in OWL and takes advantage of some speci c OWL semantics. Speci cally, the vocabulary employs the Inverse Functional Property on many of the properties connecting a person to an account. This includes foaf:aimChatID, foaf:homepage, and, most importantly for our purposes, foaf:mbox and mbox:sha1 sum.
These latter two properties connect a person to their email address, either using the email address directly or using the SHA1 hash of the address respectively. Inverse Functional Properties serve as unique identi ers; in this case, it means only one person can own a given email address. Thus, two people with the same email address can be inferred to be the same person. All the social networks that generate FOAF use the mbox:sha1 sum to identify users. The value for this property is a hash of the user's email address in the form mailto:example@example.com. Nearly all the networks provide this for every user. A few networks do not require users to provide email addresses; if no email is available, the property is not included in the FOAF output. Datasets are discussed further in section 3.
Because the mbox:sha1 sum is an inverse functional property, we can use it to merge accounts. Using an OWL reasoner, all accounts that share a common value for the mbox:sha1 sum are inferred to be the same account. 2.2
Previous research has looked at how FOAF is being used on the web. In 2005,
[
A later study in 2008 [
We are interested in nding accounts that would be merged by applying OWL
reasoning to FOAF data. Unlike previous research that studied the impact this
had on connecting social networks, we are not interested in social connections
at all. Rather, we want to know the implications that arise from identifying
multiple pro les as belonging to the same person. To do this, we used essentially
the same data as was used in [
Eleven social networking websites generate FOAF les for their users and we
used all of these networks in our research. Note that this is not just the total
number of networks we used, but all the web-based social networks with available
FOAF. LiveJournal is the largest of those, accounting for just over 75% of the
users. All of these networks were crawled in [
For each network, we gathered as many pro les as possible. Some networks - FilmTrust, Ecademy, and Advogato - provided a full list of all of their members.In the rest of the networks, a full list of members was not available, and thus we had to crawl the network. To do so, we chose several users as starting nodes and performed a breadth rst search through the network to nd all reachable members. For each user, we accessed the FOAF le, pulled URIs of their friends' FOAF les, and added those URIs to our queue. Table 1 shows the number of users in each network that we were able to use in this study.
There are almost certainly smaller components of these networks that our crawls did not reach. Also, users with no social connections would never be discovered on a crawl. However, since we are looking for multiple pro les and not examining social connections, missing pro les will not have a signi cant impact on our results. Furthermore, any applications using FOAF would need to follow the same procedures we did in this study, and thus our data set is representative of what FOAF applications would use. In the worst case, we will underestimate the number of pro les a person has, but our sample should provide representative insights into the implications of merging pro les with FOAF.
For every member we were able to include in the study, we accessed their FOAF le. For the purpose of this work, we were interested only in the member's friends and unique identi ers (given by the inverse functional properties). Thus, to save space and increase e ciency, we implemented a task-speci c OWL reasoner that considers only the FOAF inverse functional properties and foaf:knows property, and ignores the rest of the data.
Traditionally, a reasoner would not keep track of the sources of each axiom in the knowledge base. Since we are speci cally interested in how data is repeated in multiple sources, we added a provenance tracking feature to our reasoner. This maintains a record of the document where each axiom is asserted. With this data available, it is straightforward to identify on which and how many social networks a member has accounts, as well as the sources for each friendship. 3.2
Once we had collected the FOAF les as described above, we implemented a simple customized OWL reasoner that would match up pro les with the same mbox:sha1 sum. We ran this over the data and identi ed individuals with multiple accounts.These multiple accounts occurred both within and across networks, though it was much more common to have multiple accounts within a network.
Of all mbox:sha1 sums used on multiple accounts, 83.6% existed on only one network. Nearly all the rest - 16.0% - were spread between two networks. Among the mbox:sha1 sums on one network, the vast majority are on Buzznet, at 72.9%. The now defunct GreatestJournal had 15.9%, LiveJournal has 9.3%, and the rest of the networks have less than 1% of the users. This latter distribution is shown in gure 3.2.
There were 982,913 unique mbox:sha1 sum values among all the datasets. Among these, 47,563 are associated with multiple accounts. The vast majority - 77.6% (36,918) - are used on only two accounts. If we add in the 13.0% that had three associated accounts, this comprises over 90% of the total. Only 16 mbox:sha1 sums had more than 30 associated accounts. The number of accounts associated with each mbox:sha1 sum follows a power law distribution (see gure 2). The top ten mbox:sha1 sums with the most associated accounts had 350, 118, 108, 92, 62, 59, 56, 42, 40, and 39. Details about these are discussed in section 4.3 and 4.3.
In order to gain insights into why multiple accounts were associated with the same mbox:sha1 sum, we chose several subsets of accounts to examine. First, we selected the top ve email addresses with the greatest number of associated accounts. Because there are so many pro les and attributes in these accounts, we believe these users provide the clearest and most extensive picture of multiple account holding online. Then, because smaller numbers of multiple accounts were more common, we selected 40 mbox:sha1 sum values at random that had only two associate accounts and another 40 mbox:sha1 sum values with 5 associated accounts.
For each of these users, we looked at the publicly accessible personal information for each of their pro les, including name, age, gender, username, the activity on the account (last active dates, latest blog posts, etc), the content of their blog posts (when available), social networking connections, and other data that was present. We then compared and analyzed this data across the accounts held by each person to develop insights into their online personas.
If two di erent users entered the same email address, we would expect to nd very few similarities in their pro les, if any. However, if we nd ve accounts associated with one mbox:sha1 sum and the pro les of each account have the same age, gender, hometown, and astrological sign, it is highly unlikely that they belong to ve random people. The shared mbox:sha1 sum implies that the accounts belong to the same person. In our analysis of accounts, we looked for evidence to support this inference. Only when we found few or no similarities did we conclude that the accounts may belong to di erent people. 4.2
We had several theories as to why users would have multiple accounts:
1. Users opened an account and then forgot that it existed or forgot their
password. They then opened up another account to replace it. This is
easily detectable by looking at the latest activity on the account when it was
available.
2. Users create di erent accounts to compartmentalize parts of their lives. For
example, a user may have one account for personal social networking and
entertainment use, one for business use, one for religious use, etc. The goal
in this case would be to keep these parts of the user's life separate online.
3. Users create accounts for separate topics. Unlike H2 where the accounts
are used to present di erent versions of oneself to di erent audiences, this
hypothesis addresses accounts as an organizational mechanism. For example,
a blogger may have many blogs on di erent topics and create a separate
account for each. The user presents the same persona on each blog, but
separates topics through multiple accounts.
4. Users maintain completely di erent personas in di erent accounts. For
example, a user may have a pro le of a 15-year-old boy and a 50-year-old woman.
Names, locations, interests, and other personal information may also change.
The intentions behind these multiple personas vary, but this provides one of
the more interesting reasons for creating separate accounts.
5. Sybil Attacks [
We do not necessarily expect to see instances of all these theories, and we also believe we may nd other unexpected reasons behind multiple accounts. 4.3
There are insights to be found in looking at mbox:sha1 sums associated with many social networking accounts. Do these accounts actually belong to the same individual? If so, this validates the assumptions behind the FOAF model that treats an email address as a unique identi er. Understanding the purpose of these accounts also can help us understand if FOAF reasoning presents a threat to privacy that users are trying to maintain, or if it is a feature that can bene t users in maintaining their identities.
We considered the top ve most popular mbox:sha1 sums. 1. 350 Associated Accounts By far the most common value for the mbox:sha1 sum was 08445a31a78661b5c746feff39a9db6e4e2cc5cf. We found 350 pro les using this address. While it is not possible to easily unhash this address, an educated guess revealed what was happening. The sha1 hash of mailto: (i.e. an empty email address) is 08445a31a78661b5c746feff39a9db6e4e2cc5cf. Thus, when users left their address blank, systems generated this as the value for their mbox:sha1 sum. 2. 118 Associated Accounts This appears to be a test account of some sort. Ninety of the account names were consecutive integers from 2 through 90 (with 69 and 83 skipped) plus 100 and 1000. The rest of the account names were single letters or common words. The accounts are basically empty of pro le information. The only content to be found is collections of photographs in some of the accounts.
The social network of these users supports the hypothesis that they all belong to one user rather than to many di erent users. Figure 4.3(a) shows the social network for these nodes. Each red node indicates one of the 108 accounts associated with this mbox:sha1 sum while green nodes indicate accounts with di erent mbox:sha1 sum. The gure shows all the friendships of these nodes. Note that the vast majority are to other nodes that share this mbox:sha1 sum. 3. 108 Associated Accounts This was a very clear case of one user creating multiple accounts. Each pro le indicates that it belongs to a 19-year-old female living in a speci c Detroit suburb (unnamed here to protect the user's privacy). Each Social network of the 118 accounts sharing one mbox:sha1 sum. Green nodes indicate the few accounts that are in this social network but do not share the same email address.
Social network of the 92 accounts that share one address. Basically all are friends with the central node that shared the same mbox:sha1 sum. has exactly one connection to the same person, a user called\panasonicyouth", a buzznet sta member who is friends with over 575,000 users - the fast majority of people on Buzznet. There is no indication as to what the purpose is for each account - most have no content or posts. 4. 92 Associated Accounts All but three of these accounts were on the Insane Journal blogging website, and the remaining were on Dead Journal. The account names were related to pop music icons, contestants on American Idol, and characters in Disney Channel television series. The pro le photos also matched that persona (e.g. a pro le with an account name resembling Paris Hilton would also have a picture of Paris Hilton). Forty-nine of the accounts have been shut down by Insane Journal; attempts to access the pro les show the message \This journal has been deleted and purged." The remaining accounts still provide plenty of information about what is happening.
Figure 4.3(b) shows the social network of these accounts, represented as nodes. This shows all the edges for every node. Note that all the accounts are connected, and nearly all nodes have only one connection to the node at the center. That central person shares the same mbox:sha1 sum as the other accounts. None of the accounts have social connections to users with a di erent mbox:sha1 sum.
An examination of the remaining accounts reveals their intent. While most accounts have no public posts, one contains a post that explains all of the acounts. The post states that the author is a high-school student using the accounts to participate in celebrity fan ction sexual role playing games. The di erent accounts represent the many di erent personas the author represents in the games, both male and female. The games take place through private messaging, but the accounts provide a face for the personality in the game. The post explains the author's rules for engaging in a game with someone else and lists characters the student likes to represent (which correspond with the type of celebrities used in the pro le names and photos).
Thus, we can conclude that these accounts also belong to the same person who is using each as an alternative identity in role playing games. 5. 62 Associated Accounts The accounts associated with this mbox:sha1 sum do not appear to belong to the same person. The pro le characteristics vary widely. Pro les are well developed with photos, posts, and discussions. This level of e ort makes it unlikely that all the accounts are the work of one individual. It is possible that these users all entered a fake email address that we were not able to identify; the use of fake addresses was common, as we will discuss in section 5.1. 4.4
We randomly selected 40 mbox:sha1 sums that had ve associated pro les. When one or more of the pro les were private or unavailable, we threw out that mbox:sha1 sum and randomly selected another. These represented mbox:sha1 sums in the top 5% with respect to the number of associated accounts among mbox:sha1 sums with multiple accounts. We also used the same method to select 40 mbox:sha1 sums that had two associated pro les. Over 75% of the mbox:sha1 sums with multiple accounts had only two, so these are representative of the most common case.
In almost all cases, we found that the accounts associated with a given mbox:sha1 sum belonged to the same person. The pro le information was largely the same; for example, we often found the same hometown, age, astrological sign, and gender in all the pro les. This, combined with the fact that we know they used the same email address, is extremely strong evidence that all accounts represented the same person. When there were discrepancies, they were often changes in age (e.g. from under 21-years-old to 21) or hometown (e.g. from a small suburban town in one pro le to the name of the metropolitan city in another).
Of the 80 total mbox:sha1 sums, we found only four where all the accounts did not obviously belong to the same person - just over 1%.
Interestingly, most of the accounts were essentially unused. They had some basic pro le information, but no posts, no friends (beyond, occasionally, the \default" friends on a given network), no photos, and no content of any other type.When at least one account associated with a given mbox:sha1 sum has some activity, the others could possibly be mistakes, forgotten, or used for some purpose that cannot be observed. However, it was not infrequent that all the accounts had no activity. In these cases, it is unclear why users would create these accounts and not use them. They may be required to create accounts to participate in activities on the website that are not re ected on their pro le, however, there is no available evidence as to their true purpose. 5 5.1
Fake Email Addresses We found many fake email addresses in our data. We identi ed these by making up obviously fake addresses, generating the SHA1 hash, and searching for that hash in the list of mbox:sha1 sums collected from the pro les. Aside from the blank email address discussed above, table 2 shows 20 made up addresses with many associated accounts. While these are the most popular of the fake addresses we guessed at, we found many others with only one or two associated accounts.
Di erent Personas Users often change information in their pro les to represent variations on their personality or life situation. It was fairly common to see users with di erent ages but otherwise identical pro le information; we found that in 29 pro les that belonged to the same person but that had di erent ages. This was frequently a di erence between being under 21-years-old and 21 or older. We also found variations in the pro le's sexual orientation and in relationship status.
For example, Buzznet had a high population of female users in their midteens. It was not uncommon to see them create new pro les with new usernames when they started dating a boy, dedicating pictures and descriptions to their relationships, only to abandon that pro le and create a new one when the relationship ends.
Less Common Reasons { Sybil (pesudo)-Attacks - Some users created multiple accounts in order to have some accounts provide positive votes to content posted on the main account. This was only seen twice, but was a social network-based instantiation of the Sybil attack. While there was little malicious intent, this usage is similar to what would be done to circumvent rating systems in more critical environments. { Compartmentalizing - One of our hypotheses about why users had multiple accounts was due to compartmentalizing parts of their life, for example, maintaining separate accounts for their personal and professional lives. We found two mbox:sha1 sums where this was the case, but it was not a common phenomenon. { Errors - When users are not permitted to change their usernames, they register new accounts to change them. We found a few instances of users registering multiple accounts on the same day with user names that varied by one or two letters, correcting a typo. { Groups - Occasionally groups of users shared a common mbox:sha1 sum. We particularly noticed this with music groups who may be using a common address for the band while each member maintains a separate pro le. 5.2
One of our initial concerns was that people would be using multiple pro les on a social network to keep parts of their lives separate, speci cally to keep their personal life private from their professional life. However, this was not an issue that we discovered. Most pro les on the same mbox:sha1 sum were very similar and when there were variations, they were minor. The large number of empty unused pro les also lessened the concern about privacy violations.
In most cases, users were extremely open in sharing intimate details of their lives in their pro les. There were open discussions of sexual activities, drug use, and personal con icts. There were some variations between pro les in personal information (e.g. the user's sexual orientation was listed as \straight" in one pro le and \bi" in another), but it does not appear that the user was trying to hide information by using multiple pro les; rather, these instances look more like teenagers experimenting with their personas.
Even in the few cases we found where users were keeping their professional and private lives separate, the \professional" context was quite casual and included personal information.
This is not to say that FOAF aggregation will not raise privacy concerns. One only need to look to the now defunct Plink as an example. The site aggregated FOAF from many sources but was forced to shut down in October 2004 after complaints from people who did not expect their data to be present on a site they did not sign up for. Thus, even when users have only a single pro le, there are privacy concerns that arise from simply taking FOAF information that is freely available.
Independent of this general concern, however, there is little evidence that the merging of pro les will lead to information being aggregated that users intended to keep separate.
Most of the cases where Semantic Web reasoning over the FOAF pro les seems to merge accounts belonging to di erent people occur when users have fake addresses. This occurs when social networking websites do not perform an email veri cation for users to create pro les and when they allow a new user to register with an email address that's already in the system. Both of these were the case with Buzznet, resulting in the many duplicate accounts we found there. 6
In this paper, we looked at the implications if Semantic Web reasoning over FOAF data to merge multiple pro les of the same user. We were particularly interested in why users create multiple accounts, how they use them, and what challenges or bene ts FOAF o ers in dealing with this issue. To answer this question, we gathered FOAF data from all eleven social networking websites that produce FOAF les for their users, and analyzed the pro les of the ve users who had the largest number of multiple pro les as well as pro les for 40 mbox:sha1 sums with ve associated accounts and 40 mbox:sha1 sums that represented the much more common case of having only two pro les.
In all but a few cases, we found that all the accounts associated with a given mbox:sha1 sum represented the same user. We found many examples where people used fake email addresses, and this lead to many instances where pro les of di erent people were linked to the same mbox:sha1 sum.
Among the mbox:sha1 sums where the accounts belonged to the same person, we made several observations. Frequently, the pro les were mostly unused; they had only basic pro le information with no friends or posts. Users usually maintained identical information among all of their pro les. When there were discrepancies, they were usually in the user's age (perhaps to gain access to age-restricted areas of the website) or hometown. The most common reason for having multiple pro les appears to be to cultivate slightly di erent and / or evolving personas in an online environment. Merging the available FOAF information will lead to some inconsistent information in these cases. However, we found no instances where the merging of data would violate privacy that users tried to establish by separating their information into di erent accounts.