|

 

Satoss.uni.lu

An Empirical Study on User Access Control
in Online Social Networks
Faculty of Science, Technology and Shanghai Key Laboratory of Data Science University of Luxembourg Faculty of Science, Technology and Shanghai Key Laboratory of Data Science University of Luxembourg In recent years, access control in online social networks has Online social networks (OSNs) have gained a huge success attracted academia a considerable amount of attention. Pre- in the past decade. Leading players in the business includ- viously, researchers mainly studied this topic from a formal ing Facebook, Twitter and Instagram have attracted a huge perspective. On the other hand, how users actually use ac- number of users. Nowadays, OSNs have become a primary cess control in their daily social network life is left largely way for people to connect, communicate and share life mo- unexplored. This paper presents the first large-scale em- ments. For instance, every day, 500M tweets are shared on pirical study on users' access control usage on Twitter and Twitter, and Instagram users publish 60M photos1. OSNs Instagram. Based on the data of 150k users on Twitter and have brought a lot of convenience to our life, users' privacy, 280k users on Instagram collected consecutively during three on the other hand, has become a major concern due to the months in New York, we have conducted both static and dy- large amount of personal data shared online. Previously, namic analysis on users' access control usage. Our findings researchers showed that a user's personal information can include: female users, young users and Asian users are more be inferred through statuses [18] and locations [25] that she concerned about their privacy; users who enable access con- shared in OSNs.
trol setting are less active and have smaller online social To mitigate users' privacy concern, major OSNs have de- circles; global events and important festivals can influence ployed access control schemes to delegate the power to users users to change their access control setting. Furthermore, themselves to control who can view their information. For we exploit machine learning classifiers to perform an access example, Facebook provides a fine-grained access control control setting prediction. Through experiments, the pre- scheme which enables users to apply different policies on dictor achieves a fair performance with the AUC equals to each post they publish. Twitter and Instagram, on the other 0.70, indicating whether a user enables her access control hand, provide a much simpler scheme. A Twitter or Insta- setting or not can be predicted to a certain extent.
gram user could enable her access control setting such thatstrangers cannot have access to all detailed contents in heraccount, except for her profile picture, number of friends and Categories and Subject Descriptors number of online posts. To study and further improve access K.6.5 [Management of Computing and Information control in OSNs, academia have conducted many research, Systems]: Security and Protection; H.2.8 [Database Man- most of which take either formal or logical approaches. For agement]: Database Applications—Data mining instance, researchers have modeled access control with hy-brid logic [11] and semantic web technology [3].
other hand, understanding how users exploit access control in their daily life is essential to improve access control in Online social networks; access control; empirical analysis; OSNs. Much to our surprise, this is left largely unexplored.
In this paper, we perform a large-scale empirical study on access control usage of Twitter and Instagram users inNew York. To the best of our knowledge, this is the first Permission to make digital or hard copies of all or part of this work for personal orclassroom use is granted without fee provided that copies are not made or distributed work on analyzing users' access control on Twitter and In- for profit or commercial advantage and that copies bear this notice and the full cita- stagram. We collect data of 150k Twitter users and 280k In- tion on the first page. Copyrights for components of this work owned by others than stagram users continuously within three months and study ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or re-publish, to post on servers or to redistribute to lists, requires prior specific permission their access control usage from both static and dynamic and/or a fee. Request permissions from [email protected].
point of view. Especially, the dynamic analysis is conducted SACMAT'16, June 05-08, 2016, Shanghai, China on a daily base instead of a yearly base as done in previ- 2016 ACM. ISBN 978-1-4503-3802-8/16/06. . $15.00 ous works [8, 21]. This allows us to understand in depth Access control schemes on these three OSNs are differ- how users exploit access control in their daily OSN life. Our ent as well. Facebook deploys a fine-grained access control contributions in this paper can be summarized as follows.
scheme for users to control who can view their resources.
This scheme is on a per-resource base, i.e., a user can de- • We perform a static analysis on New York users' ac- fine a specific access control policy for each of her photos cess control usage and find that female users and young and statuses. In addition, Facebook also introduces a func- users are more likely to enable their access control set- tion, namely friend list to help users categorize their friends ting. Moreover, users who enable their access control into different lists, e.g., colleagues and family, and the orga- setting tend to have smaller online social circles, but nized friend lists can then be directly used in a user's access are more willing to conduct social activities in the of- control policy which improves its access control scheme's us- fline world represented by location check-ins.
ability. Different from Facebook, Twitter4 and Instagram5provide users with a much simpler access control scheme.
• We conduct a dynamic analysis on users' access con- On Twitter and Instagram, users can only choose whether trol usage based on the three-month consecutive data.
to enable their access control setting, i.e., protect their ac- We find that a considerable amount of users change count or not. Once a user enables her access control setting, their access control setting frequently and there are others who are not the user's approved followers cannot view more users (especially female users and young users) any of her information except for her profile photo, number enabling their access control setting than disabling it.
of followers/followees and number of posts. In the follow- When users disable the access control setting, they ing analysis, we refer users who enable their access control tend to become less active online and delete some of setting as private users while others as public users.
their followers. Interestingly, we also find that impor- To improve access control in OSNs, one important per- tant festivals and events cause more users to disable spective is to understand how users apply their access con- access control setting.
trol in their OSN life. Several previous works [12, 8, 21, 13]have focused on the access control usage on Facebook. How- • We apply machine learning techniques to conduct a ever, to the best of our knowledge, there do not exist works prediction on whether a user would enable her access focusing on Twitter and Instagram.
As discussed above, control setting or not. By combining users' online be- these two OSNs deploy different access control schemes from havior such as the number of followers, together with Therefore, it is important and meaningful to user demographics, our prediction experiments achieve understand how users apply their access control setting on a fair result in which the AUC (area under the ROC Twitter and Instagram to protect their privacy.
curve) equals to 0.70. This indicates a user's accesscontrol setting can be predicted to a certain degree.
In this paper, we collect the access control usage data The rest of the paper is organized as follows. Section 2 of New York users on Twitter and Instagram. Even though introduces background information of Twitter and Insta- the dataset of New York users is not a random sample of the gram's access control schemes as well as the dataset used global population, due to the diversity of New York users [8], for our study. Section 3 and Section 4 present static and dy- we believe that our analysis should be indicative enough to namic analysis on users' access control usage, respectively.
reflect users' access control usage in general.
Section 5 performs an access control setting prediction using To identify users in New York, we leverage check-ins (user- machine learning techniques. Section 6 discusses limitations shared location information) on the two OSNs. Nowadays, of this paper. Section 7 summarises related work and Sec- many people use their OSN services on mobile devices, e.g., tion 8 concludes the paper with some future works.
80% of Twitter's active users are on mobile6. To adapt tothis trend, major OSNs add new functionalities to their mo- BACKGROUND AND DATASET bile versions, one of which is location sharing, namely check-in, through mobiles' GPS sensors. It is quite common for Access Control in OSNs users to share a photo together with the location where thephoto is taken. By exploiting check-ins to identify users in Facebook, Twitter and Instagram are among the most New York, we can ensure accurate results from our analysis.
popular OSNs at the moment. By September 2015, Face- Moreover, it allows us to compare public and private users' book has around 1.5 billion monthly active users with 83.5% mobility behaviors as well (see Section 3).
of its users are outside the US and Canada2, while Twitter To obtain users' check-ins, we first define a geo-coordinate and Instagram have 316 million and 400 million monthly bounding box covering New York region and then exploit active users respectively. Besides the difference in size, the Twitter [16] streaming API7 and Instagram REST API8 to three OSNs are also appealing to different demographics collect users' check-ins respectively. To make sure that the and usage. Facebook is a general purpose OSN3 with users users are locals in New York rather than visitors, we only distributed more evenly to diverse ages, races and genders; keep those users with more than 10 check-ins.
Twitter on the other hand is largely treated as a news source, depicts a sample of check-ins in New York on Instagram.
also its percentage of users with high education and income After identifying the users in New York, we use Twitter is higher than those of the other two OSNs; Instagram is a platform for users to share their life styles and its users are more skewed to young people.
Table 1: Summary of the conducted analysis on Twitter and Instagram.
a) Gender
s
er
us 4000
Number 2000
Figure 1: Check-ins in New York on Instagram.
REST API9 and Instagram REST API to extract users' ac-cess control setting together with some general information Figure 2: Gender, race and age distributions of New such as number of followers/followees and number of posts, York users on Twitter and Instagram.
on a daily basis for nearly three months, from October 15th,2015 until January 12th, 2016. We regard accounts withmore than 2,000 followers as celebrities and those whose fol- ternational competitions, it has also been exploited in other lowers are 1,000 more than followees as business accounts, works for detecting users' demographics, such as [20, 19].
and remove them from the dataset.
Figure 2 depicts gender, race and age distributions of our For static analysis, we focus on the data collected on users on Twitter and Instagram. As we can see, there are November 12th with 175,202 users for Twiter and 292,406 more female users than male users on Twitter and Insta- users for Instagram10. For dynamic analysis, we focus on gram in New York, and the proportion of female is much users that appear in our dataset everyday. In the end, we higher on Instagram. In addition, as mentioned before that get 155,387 Twitter users and 282,066 Instagram users11.
Instagram attracts more young users than Twitter, thus its Note that when we exploit API to extract a private user's users' age distribution is skewed to younger ages than that information, Twitter allows us to access the user's profile of Twitter users.
photo, number of followers/followees, and number of postswhile Instagram forbids all the access. Since we get users' demographics through analyzing their profile photos, and In this section, we perform static analysis on users' access quantify their online behavior through their numbers of fol- control usage. We start by checking the percentage of pri- lowers/followees, and numbers of posts, we cannot conduct vate users in our dataset, then analyze the relation between analysis related to those information on Instagram users.
users' demographics and their access control usage. Users' Table 1 lists the analysis we perform on the two OSNs.
online and offline behavior is discussed in the end. Here, a Users' demographic information is another important as- user's online behavior is quantified by her number of posts pect of our analysis. To get users' demographics, we resort to and followers/followees in OSNs, while offline behaviors are Face++12, a state-of-the-art facial recognition service that quantified by her mobility, i.e., check-ins. As mentioned in detects a user's gender, race (Asian, White, African Ameri- Section 2, we have no access to the demographic information can) and age information from her profile photo. Face++ is and online bahaviors of private users on Instagram, thus we based on deep learning techniques and has won several in- cannot perform static analysis on Instagram users' demo-graphics and online behaviors.
10We have analyzed data collected on other dates and the General Statistic analysis results are similar.
As shown in Table 2, the general percentages of private Some users might delete their accounts or get suspended during the three months, thus the number of users for dy- users in our dataset are 5.22% for Twitter and 11.92% for namic analysis is slightly smaller than the number of users Instagram. This indicates that Instagram users pay more used for static analysis.
attention to their privacy than Twitter users. The reason could be the different purposes of using the two OSNs (see Section 2): Twitter is treated as a news spreading medium, thus its users are less likely to share personal sensitive in- formation; Instagram, on the other hand, is a photo-sharing OSN and photos can contain personal sensitive information.
Table 2: Statistics of public and private users.
% of Private Users % of Public Users % of Public Users % of Private Users a) Gender
Private User Percentage Figure 3: Demographic distributions of Twitter users.
So far there does not exist official data from Twitter and Instagram on their private users' percentages. Cha et al. [5] claim that the percentage of private Twitter users is more than 7% which is close to our observation. The slight differ- ence can be due to the sampling methodologies. The dataset in [5] is sampled through randomly picking user ids, while our dataset focuses on users in New York. On the other hand, the percentage of private users on Instagram is un- % of private users % of public users % of public users % of private users a) Gender
clear from the literature. We emphasize that the focus ofthis paper is to understand how users exploit their accesscontrol in real life, the general percentages of private users Figure 4: Demographic distributions of all users and on Twitter and Instagram are certainly interesting but left users whose ages are below 10 on Twitter.
as future work.
We further analyze users under 10 through their race and gender. Compared with Figure 3, Figure 4 shows that the Observation 1: Access control usage is different amongusers with different genders, races and ages. Female users, percentages of private users have increased for both gen- young users and Asian users are more likely to enable their ders and all races. The percentage of female private users access control setting than others.
increases 1.33 percents to 8.24%, while that of male usersincreases 1.84 percents to 5.99%. Furthermore, Asian users Gender. We calculate private users' percentages of male remain to have the most private users, with the percent- and female users respectively, and find out that more female age of private users increases about 2.1 percents to 8.3%.
users enable their access control setting than male users. As Those two results coincide with the results in our analysis we can see from Figure 3, 4.15% of male Twitter users enable above that female users and Asian users are more concerned their access control setting while the corresponding rate of about their privacy. More importantly, we can see that users female users is 6.91%.
who use children's photos in their profiles tend to be more Race. Among people of three races in New York (see Fig- ure 3), the private users' percentage of Asian users is the Profile pictures.
As introduced in Section 2.2, we ex- highest (6.20%) followed by White users (5.60%). African tract a user's demographic information through recognizing American users, on the other hand, have the lowest percent- her profile photo with Face++. Therefore, if a user uses age (5.22%). One possible explanation could be the culture non-human pictures, such as a cat, in the profile, we can- difference: Asian people are considered more conservative not get her demographics14. Private users' percentages of than White and African American people in general13.
Twitter users with and without demographics are listed in Age. Figure 3 shows that for all Twitter users who are older Table 2. Among 107,413 users with demographics, 6,066 than 10, the percentages of private users are decreasing when (5.65%) users are private, while for 67,789 users without de- the age grows.
This trend is especially notable for users mographics, only 3,079 (4.54%) of them are private. This aged from 20 to 40, which indicates younger people are more indicates that users without demographics (not using human concerned about their privacy than people of other ages.
photos) on Twitter care less about their privacy than those Interestingly, the private users' percentages of users under who use human pictures. The reason might be that using 10 years old (children) are high as well. Since children under fake profile pictures makes users feel secure.
10 are less likely to be frequent Twitter users, we conjecturethat these users use children's photos in their profiles.
14By manually checking 100 users without demographics in our dataset, we find that more than 90% of these users use non-human pictures in their profiles.
the other hand, private users following less people is an in-teresting observation. This suggests that private users tend to filter not only their followers, but also followees to ensure their social circles to be less chaos.
Observation 3: Private users are more socially active than public users in the offline world.
The mobility data we get from Twitter and Instagram can be a good reflection of New York users' offline life. Ourmobility dataset is composed of users' check-ins, and each check-in of a user tells us when and where the user is. In thefollowing, we conduct our analysis from these two aspects.
Time. Figure 6 depicts the distributions of users' check-intime on a daily base on Twitter and Instagram. Despite thedifferent distribution curves (Twitter users are more active at late night and early morning), we can observe an agree- ment between the two OSNs: compared to public users, pri-vate users are more active at night. As most offline social Figure 5: Users' distributions of posted tweets (top- activities happen at night rather than working hours, this left), favored tweets (top-right), followers (bottom- indicates that private users are more socially active in the left) and followees (bottom-right).
offline world.
Locations. Due to the different designs of the two OSNs' APIs, we can extract the category information of each lo-cation for Instagram while not for Twitter. Here, location Observation 2: Private users share more contents but have category information on Instagram is from Foursquare, a smaller online social circles than public users.
popular location-based social network, in which different lo- Four metrics are exploited to quantify a Twitter user's cation categories are organized into a tree structure15. In online behavior, including the number of posted tweets, the this paper, we take the first layer of the category tree (nine number of favorites, the number of followers and the number categories) to label each location, including entertainment, of followees. The former two metrics can be used to evaluate university, food, nightlife, outdoor, professional, residence, the active level of each user on Twitter, while the latter two store and transportation.
represent the size of each user's social circle.
Figure 7 depicts distributions of public and private users' We use box plots [23] to visualize the distributions of four check-ins over different location categories. It shows that metrics for private and public users respectively (see Fig- private users have more check-ins at food and nightlife places.
ure 5). Box plot, i.e., box and whisker diagram, is a stan- Since many offline social activities happen at these two types dardized way of displaying data distribution based on the of places, this further confirms that private users are more five number summary: minimum, first quartile (25%), me- active in the offline world.
dian, third quartile (75%), and maximum.
From Figure 5, we observe that, compared with public users, private users have published and favored more tweets.
Our static analysis focuses on three aspects including de- On average, private users have posted 8,864.07 tweets and mographics, online behavior and offline behavior. We have favored 3,380.43 tweets, while public users posted 7,550.96 tweets and favored 2,277.58 tweets. There are two possibleexplanations for this result: • Female users, young users and Asian users are more concerned about their privacy than others; • Private users publish more tweets, i.e., to express them- selves, since they are aware that their privacy is guar- • Private users publish more posts than public users, but anteed to a certain extent; have smaller online social circles; • Users who have used Twitter for a longer period of • In the offline world, private users are more socially time are more likely to become private since they are active than public users.
more aware of the privacy threats. Meanwhile, theirlonger Twitter-ages result in more tweets.
In Section 4, we perform further analysis on this from a After the static analysis, in this section we study New dynamic point of view.
York users' access control usage from a dynamic perspec- Public users have more followers and followees than pri- tive. Questions we attempt to answer include: how many vate users (see Figure 5). On average, public users have users have changed their access control setting; what is the 423.26 followers and 431.73 followees, while private users changing trend; who are these users; what is the correlation have 329.37 followers and 355.17 followees. This indicates between users' changing of access control and other factors that private users have much smaller social circles. Fewer such as online behavior and global events? followers may due to Twitter's access control scheme sinceprivate users have to give approvals to their followers. On check-in distribution check-in distribution Figure 6: Check-in distribution over time on Twitter (left) and Instagram (right).
Table 3: Statistics of users' changing frequency.
check-in distribution 0.05 Figure 7: Check-in distributions over location cat- egories on Instagram, each category is represented by its first letter, e.g., F stands for food.
We start by checking the general statistics of users who change access control setting, then focus on these users' de- mographics. Next, the correlation between access control changes and users' online behavior is analyzed. In the end, we study the influence from global events and festivals on users' decisions of changing access control.
General Statistics Observation 4: A considerable amount of users' access that users' privacy concerns are increasing day by day. Note control usage is dynamic, i.e., they change their access con- that similar reslts are obtained for New York [8] and Pitts- trol setting from time to time. There are more users chang- burgh [21] users on Facebook.
ing their access control setting from public to private thanfrom private to public.
Changing frequency. A considerable amount of users inour dataset have changed their access control setting during Observation 5: Female users and young users change ac-cess control setting more frequently and have a faster chang- the three months. On Twitter, 7,590 (5.21% of total Twit- ing trend from public to private than others. White users ter users) users have changed their access control setting, change access control setting least frequently and their chang- while the proportion on Instagram is much higher (19.95% ing trend from public to private remains the slowest.
of 56,261 users).
Table 3 further presents the statistics of times that users Changing frequency and demographics. The statistics have changed their access control setting. Among all the of both Twitter and Instagram16 users' changing frequency Twitter users who have changed their access control, 54.44% w.r.t. demographics is presented in Table 4, 6.52% of female of users have changed more than once. On the other hand, users and 3.66% of male users on Twitter, and 19.20% of Instagram users seem to be more indeterminate on access female users and 13.92% of male users on Instagram have control usage: 69.09% of them have changed more than once changed their access control setting.
In addition, female during the three months. Moreover, 553 Instagram users users change access control more frequently than male users.
have even changed more than 15 times.
Especially on Instagram, the average changing times for fe-male users is 3.60, while it is 2.93 for male users.
Changing trend. From the dataset, we have observed an On Twitter, Asian users have the highest proportion of ac- increasing trend of users enabling their access control set- cess control changing (6.11%), while African American users ting during the three months, i.e., more users change from have the most frequent changing times, i.e., 2.40 times on av- public to private than changing from private to public. OnOctober 14th 2015, 4.89% of Twitter users and 9.36% of 16As some public users with demographics on Instagram Instagram users in our dataset are private, while on Jan- change their access control setting to private during the uary 12th 2016, the percentages have increased to 5.62% three months, we can still study Instagram users' changing on Twitter and 14.20% on Instagram. This result indicates frequency and trend w.r.t. demographics here.
Table 4: Statistics of users' changing frequency on Twitter and Instagram w.r.t. demographics.
users changed (%) average changed times users changed once (%) users changed (%) average changed times users changed once (%) erage. On the other hand, 20.63% of African American usershave changed their access control setting on Instagram, but Asian users have the most changing times, 3.84 on average.
On both Twitter and Instagram, White users are the most determinate about their access control setting.
We discretize age into four bins and study users' chang- ing frequency w.r.t. each age bin.
On both Twitter and Instagram, younger users change their access control setting more frequently. For instance, 18.09% Instagram users be- tween 11 and 30 years old have changed at least once and the average changing times is 3.74. While for users between31 and 45 years old, the two number is 13.14% and 2.82 respectively. In addition, users under 10 is the group with the highest number of users who change their access control setting freqently, this is consistent with our previous analy- sis that users under 10, i.e., users using children's photos in their profiles are more concerned about their privacy.
Changing trend and demographics. We further study the changing trend of users w.r.t. demographics. As shown in Figure 8a), female private users' percentage grows faster than that of male users. On Twitter, the percentage of pri- vate female users increases 0.94%, while that of private male Below 10(Twitter) Below 30(Twitter) users is 0.72%. This trend is more obvious on Instagram, Below 45(Twitter) private female users' percentage increases nearly 10% while Above 46(Twitter) Below 10(Instagram) private male users' percentage increases about 8%.
Below 30(Instagram) Trends of enabling access control setting by users of differ- Below 45(Instagram) Above 46(Instagram) ent races are exhibited in Fig. 8b). On Twitter, the propor- tion of private African American users increases the slowest, while on Instagram, it becomes the fastest. We believe it is caused by the different purposes of the two OSNs.
The changing trend for users of different age (bin) is plot- Figure 8: Changing trends of proportions of private ted in Fig. 8c). On both Twitter and Instagram, the private users based on their demographics in our dataset.
users' percentage of younger users increases faster than older As we cannot get private users' demographics from users. This accords with the result in Section 3 that young Instagram in the beginning, thus the beginning pro- users are more concerned about their privacy.
portion for Instagram users w.r.t. different demo- graphics is approaching 0.
Observation 6: In general, users being private through allthe three months and users changing from public to private ing 1) new tweets; 2) new favorites; 3) new followers; and 4) tend to be less active in publishing new contents. Besides, new followees, added daily17.
these users barely establish new relationships with others, The comparisons between the constantly-public users and and their followers become fewer. Moreover, topics of users' constantly-private users w.r.t. four metrics of dynamic on- posts on both OSNs are more (less) personal/sensitive when line behavior are shown in Figure 9. Recall the observation changing from public (private) to private (public).
in Section 3 that private users have more tweets (and fa- Statistics of online behavior. We first refer users staying vored tweets) than public users in general. Here, we find private (public) within the three month as constantly-private 17Different from demographics, users' online behaviors are (constantly-public) users, users who have changed their ac- dynamic, for instance, number of followers may vary every- cess control setting are named inconstant users. Based on day. Thus, we cannot apply the same method for demo- users' online behavior presented in Section 3, we design four graphics to analyze Instagram users' dynamic online behav- metrics to evaluate users' dynamic online behavior, includ- ior in this section.
Table 5: Statistics of inconstant users' dynamic on- line behaviors.
Public to private Private to public California Gun Shot Major Baseball League's b) Social Network Figure 9: Constantly-public and constantly-private users daily added new tweets, favored tweets, fol- lowers and followees.
that, on the contrary, constantly-private users have less new California Gun Shot and favored tweets everyday than constantly-public users on average. In Figure 9a), both curves representing constantly- Major Baseball League's private users are below the ones for constantly-public users.
In Section 3, the two explanations why private users hav- ing more tweets than public users include: users are morecomfortable to express themselves in the private context; Figure 10: Differences between daily new private private users have longer Twitter-ages, thus having more users and daily new public users.
tweets. The result in Figure 9a) gives a strong support forthe second explanation, i.e., longer Twitter-age is the reasonwhy private users have more Tweets than public users.
start to query each inconstant user's posts through the cor- We also find that constantly-private users barely establish responding API one day before and after she changed her new links with others, i.e., their average amount of daily new access control setting, then aggregate posts of each user to- followers and followees are very close to 0, while constantly- gether as one document. Punctuations and stop words are public users often have new followers and followees every filtered out during the process. We then organize all the day. This suggests that private users are more careful on documents into a corpus, and remove words that appear in choosing their followers and followees (similar to the sum- less than 20 documents and more than 70% of the docu- mary in Section 3).
ments [24]. Note that for users who change from public to Inconstant users' dynamic online behavior statistics are private, we cannot get their published posts on both Twitter presented in Table 5. It appears that users changing from and Instagram. However, as users frequently change their public to private have fewer newly posted and favored tweets access control setting and some private users become pub- than those changing from private to public. Moreover, in- lic on the day we collect their published posts (January 21, constant users changing from public to private are reducing 2016), we are able to extract topics from their posts when their followers and followees. In addition, more followers are they change from public to private.
deleted than followees (-0.15 vs. -0.04), which indicates that Table 6 lists the top 3 topics for Twitter and Instagram users' one purpose of enabling access control is, to some ex- when users change their access control setting. We observe tent, to protect themselves from being viewed by someone that when users are public, their topics are not privacy- from whom they are hiding sensitive information. In an- sensitive, for instance, "happy, years, new" published during other way, users are more concerned about privacy leakage the New Year on Twitter and "follow, keep, coming" repre- through who follows them than who they follow. This result senting the popular hashtags on Instagram. On the other reflects some fundamental differences between follower and hand, when users enable their access control setting, their followee relations on Twitter.
topics become more private, such as "family" on Twitter and User topics. Next, we analyze posts (tweets for Twitter "missing" on Instagram.
and captions of photos for Instagram) that users publish Global Events and Festivals before and after they change access control setting and checkwhether topics of users' posts have changed.
Observation 7: Global events and festivals cause more users We exploit a classical topic modeling algorithm in the nat- to change access control setting from private to public.
ural language processing field, namely Latent Dirichlet Al- As stated before, the trend for users to enable their access location (LDA) [1] to detect topics from users' posts. We control setting is increasing, thus there should be more users • Many users change their access control setting from Table 6: Topics of users' posts one day before and time to time. Instagram users change more often than after changing their access control settings.
Twitter users. More users change from public to pri- Public to private Private to public vate, showing that users become more concerned about their privacy day by day.
• Female users and young users change their access con- trol setting more frequently and their changing trend from public to private is faster than others. Asian and African American users behave differently on Twitter and Instagram, while White users' changing behavior is the least active on both OSNs.
• Constantly-private users are less active than constantly- public users in terms of published posts and new fol- lowers/followees. When users change from public to Public to private Private to public private, they publish less tweets than users changing from private to public, and delete their followers, their posts' topics are more privacy-sensitive than before.
• Global events and festivals cause more users to change access control setting from private to public.
ACCESS CONTROL PREDICTION After analyzing users' access control usage, in this section we investigate whether it is possible to predict a user's ac- cess control setting. Being able to predict a user's access control setting opens up opportunities for appealing appli-cations. For instance, OSNs can automatically assign accesscontrol setting to their users for better privacy protection;government can develop a privacy advisor to remind users changing from public to private than from private to public of their privacy leakage. Our prediction is based on users' every day. However, when plotting the difference between static information, by only using a user's information listed daily new private users and public users (the number of new on the OSN page, we aim to predict whether the user should private user subtracts the number of new public users), we enable access control setting or not.
have found several interesting dates on which many more We model access control prediction as a binary classifi- users changed from private to public than from public to cation problem, and intend to solve the problem with ma- private (see Figure 10).
chine learning classifiers.
We label private users as posi- On three important festivals in the US, i.e., Thanksgiving tive cases while public users as negative cases. For features (November 26th, 2015), Christmas (December 25th, 2015) used in classification, two models are constructed, namely and New Year's Day (January 1st, 2016), more users disable Model1 and Model2. Model1 exploits users' static online be- their access control setting on both Twitter and Instagram.
havior including number of followers/followees, number of This indicates that on holidays, users are more open and less tweets and number of favorites as features for classification.
concerned about their privacy for the purpose of meeting Model2 combines the features of Model1 with demographics.
new people and expressing gratitude.
In demographics, there are two categorical variables includ- In addition, we find that some global events might cause ing gender and race, we change them into dummy variables more people to become public in OSNs as well. For instance, for classification. Three machine learning classifiers, i.e., lo- more users become public on November 13, 2015 (Paris ter- gistic regression, random forest and gradient boosting, have rorist attack) and on December 3rd, 2015 (California gun been used to conduct prediction. ROC (Receiver operating shot case). This is probably because users are more willing characteristic) curve and AUC (area under the ROC curve) to express their opinions when such events take place.
are used as evaluation metrics.
There also exists an obvious drop on November 1, 2015 on both OSNs, we believe this is due to the final match ofthe Major Baseball League's champion series between New Table 7: AUC of prediction.
York Mets and Kansas City Royals held in New York. Even though New York Mets lost the championship on that day,there are still New York users becoming public to commu- logistic regression nicate with other baseball fans on Twitter and Instagram.
gradient boosting In this section, we study dynamic usage of users' access Table 7 lists the AUC for two models under each clas- control in OSNs and have observed the following.
Our best prediction result (gradient boosting and policies are defined by semantic web technologies. The au-thors of in [9, 2] propose to use hybrid logic as the policy lan- guage, this logic has been demonstrated quite powerful and later be used in several works including [22, 7, 14, 17, 6, 15].
Compared to the formal perspective, not many works fo- Logistic regression Logistic regression cus on the empirical perspective of access control in OSNs.
True Postive Rate 0.2 True Postive Rate 0.2 Gradient boosting Gradient boosting Existing works include [12, 8, 21, 13]. Compared to these works, this paper has the following advantages: False Postive Rate False Postive Rate Figure 11: ROC curves for We perform dynamic analysis on users' access control Model1 (left) and Model2 (right) w.r.t. different classifiers.
usage within three consecutive months, which allowsus to study users' change from a daily perspective andprovide with more insightful conclusion on users daily Model2) is fair, the AUC equals to 0.7018, which indicates online activities. On the other hand, the dynamic anal- that users' access control setting can be predicted to a cer- ysis in [8, 21] is yearly-based, which can only provide a tain extent. Model2 achieves a better result than Model1 in- general trend of access control usage. For instance, the dicating demographics' usefulness on separating public and authors of [21] study users' access control setting once private users. Figure 11 further depicts the ROC curves for a year from 2005 to 2011 and discover that more and Model1 and Model2, respectively.
more Facebook users in Pittsburgh enable their access As we cannot get private users' demographics and online control setting every year.
behavior information from Instagram's API (see Section 2),we only focus on Twitter users for access control prediction.
• This paper conducts a much more comprehensive study than previous ones ranging from users' demographicsto online behavior. Besides, we are the first to analyze the relation between access control and users' offline In this section, we discuss a few limitations in our study.
behaviors (mobility information), topics of published Dataset. The current work focus on New York users' access texts and global events.
control usage. Even though the user sample is large (morethan 150k users for Twitter and and more than 280k users • This paper is the first to show that it is possible to for Instagram), there still exist some region bias in our anal- use users' information to predict their access control ysis. Meanwhile, knowing users being in New York allows us setting to a certain extent. This result can potentially to conduct some more interesting analysis, such as users' of- lead to promising applications such as automatic ac- fline behaviors (see Section 3) as well as a base ball match's cess control enforcement and privacy advisor.
influences on users' access control changing (see Section 4).
• Our user sample is bigger than most of the previous Social relation and access control. Section 4 concludes works. We have more than 150k users for Twitter and that when a user changes her access control setting from more than 280k users for Instagram while the dataset public to private, the user is more likely to reduce her num- in [12] focuses on 200 users and the one in [21, 13] has ber of followers. However, we haven't conducted the detailed around 1,000 users. On the other hand, the authors analysis on who are these users being deleted. One obstacle of [8] use a bigger sample than us (1.4 million New for this analysis is the restriction of the API: Twitter allows York users on Facebook).
much less access to their users' social networks19, while In-stagram provides only a small number of followers/followees Besides the above advantages, all of [12, 8, 21, 13] only of a user each time20.
focus on Facebook while, to the best of our knowledge, thisis the first work to analyze users' access control usage on Twitter and Instagram. As stated in Section 2, Twitter andInstagram have different types of users and functions com- Access control in OSNs has attracted academia a consid- pared to Facebook, thus it is very interesting and meaningful erable amount of attention during the past decade. Many to study their users' access control usage.
researchers have focused on modeling access control schemesin OSNs from a formal or logical perspective. Carminati etal. [4] propose three regulations for access control scheme in CONCLUSION AND FUTURE WORK OSNs, including social relation, distance in social network We have conducted the first large-scale empirical study on as well as trust level. The authors of [10] describe a two- users' access control usage on Twitter and Instagram. Our stage access control where, to access a resource of a certain analysis focused on both static and dynamic perspectives.
user, one has to be able to reach that user in the social net- We further demonstrated that users' access control setting work and then requests the access. Besides modeling access can be predicted to a certain extent.
control schemes, researchers have also proposed methods to For the future work, we plan to conduct analysis on other precisely define access control policies. In [3], access control cities to check whether culture differences play a role in 18AUC is not sensitive to label imbalance problem and AUC users' access control usage. Our access control prediction is for random guessing is around 0.5.
only based on users' static information, we plan to explore users' dynamic information to predict whether a user will change her access control setting on a certain day by using more sophisticated features in machine learning classifiers.
[14] J. Pang and Y. Zhang.
A new access control scheme for Facebook-style social networks. In Proc. 9th Conference on Minyue Ni and Weili Han are supported by NSFC (Grant No.
Availability, Reliability and Security (ARES), pages 1–10.
61572136). We thank anonymous reviewers for their comments.
IEEE CS, 2014.
[15] J. Pang and Y. Zhang. Cryptographic protocols for enforcing relationship-based access control policies. In Proc. 39th An- [1] D. Blei, A. Ng, and M. Jordan. Latent dirichlet allocation.
nual IEEE Computers, Software & Applications Conference Journal of Machine Learning Research, 3:993–1022, 2003.
(COMPSAC), pages 484–493. IEEE CS, 2015.
[2] G. Bruns, P. W. L. Fong, I. Siahaan, and M. Huth.
[16] J. Pang and Y. Zhang. Location prediction: communities Relationship-based access control: its expression and en- speak louder than friends. In Proc. 3rd ACM on Conference forcement through hybrid logic. In Proc. 2nd ACM Con- on Online Social Networks (COSN), pages 161–171. ACM, ference on Data and Application Security and Privacy (CO- DASPY), pages 117–124. ACM, 2012.
[17] J. Pang and Y. Zhang.
A new access control scheme for [3] B. Carminati, E. Ferrari, R. Heatherly, M. Kantarcioglu, Facebook-style social networks.
Computers & Security, and B. Thuraisingham. A semantic web based framework 54:44–59, 2015.
for social network access control. In Proc. 14th ACM Sym-posium on Access Control Models and Technologies (SAC- [18] M. J. Paul and M. Dredze. You are what you tweet: Analyz- MAT), pages 177–186. ACM, 2009.
ing twitter for public health. In Proc. 5th AAAI Conferenceon Weblogs and Social Media (ICWSM), pages 265–272. The [4] B. Carminati, E. Ferrari, and A. Perego. Rule-based access AAAI Press, 2011.
control for social networks. In Proc. IFIP WG 2.12 and 2.14 [19] M. Redi, D. Quercia, L. Graham, and S. Gosling. Like party- Semantic Web Workshop (OTM), volume 4278 of LNCS, ing? your face says it all. Predicting the ambiance of places pages 1734–1744. Springer, 2006.
with profile pictures. In Proc. 9th AAAI Conference on We-blogs and Social Media (ICWSM), pages 347–356. The AAAI [5] M. Cha, H. Haddadi, and F. B. K. P. Gummadi. Measuring Press, 2015.
user influence in Twitter: The million follower fallacy. InProc. 4th AAAI Conference on Weblogs and Social Media [20] F. Souza, D. de Las Casas, V. Flores, S. Youn, M. Cha, (ICWSM), pages 10–17. The AAAI Press, 2010.
D. Quercia, and V. Almeida. Dawn of the selfie era: Thewhos, wheres, and hows of selfies on Instagram. In Proc. 3rd [6] M. Cramer, J. Pang, and Y. Zhang. A logical approach to ACM on Conference on Online Social Networks (COSN), restricting access in online social networks. In Proc. 20th pages 221–231. ACM, 2015.
ACM Symposium on Access Control Models and Technolo-gies (SACMAT), pages 75–86. ACM, 2015.
[21] F. Stutzman, R. Gross, and A. Acquisti. Silent listeners: The evolution of privacy and disclosure on Facebook. Journal of [7] J. Crampton and J. Sellwood. Path conditions and principal Privacy and Confidentiality, 4(2):2, 2013.
matching: a new approach to access control. In Proc. 19thACM Symposium on Access Control Models and Technolo- [22] E. Tarameshloo, P. W. L. Fong, and P. Mohassel. On pro- gies (SACMAT), pages 187–198. ACM, 2014.
tection in federated social computing systems. In Proc. 4thACM Conference on Data and Application Security and Pri- [8] R. Dey, Z. Jelveh, and K. Ross. Facebook users have be- vacy (CODASPY), pages 75–86. ACM, 2014.
come much more private: A large-scale study.
2012 IEEE International Conference on Pervasive Comput- [23] J. W. Tukey. Exploratory Data Analysis. Pearson, 1977.
ing and Communications Workshops, pages 346–352. IEEE,2012.
[24] W. X. Zhao, J. Jiang, J. Weng, J. He, E.-P. Lim, H. Yan, and X. Li. Comparing Twitter and traditional media us- [9] P. W. L. Fong.
Preventing sybil attacks by privilege at- ing topic models.
In Proc. 33rd European Conference on tenuation: a design principle for social network systems.
IR Research (ECIR), volume 6611 of LNCS, pages 338–349.
In Proc. 32nd IEEE Symposium on Security and Privacy Springer, 2011.
(S&P), pages 263–278. IEEE CS, 2011.
[25] Y. Zhong, N. J. Yuan, W. Zhong, F. Zhang, and X. Xie. You [10] P. W. L. Fong, M. M. Anwar, and Z. Zhao. A privacy preser- are where you go: Inferring demographic attributes from vation model for Facebook-style social network systems. In location check-ins. In Proc. 8th ACM International Con- Proc. 14th European Symposium on Research in Computer ference on Web Search and Data Mining (WSDM), pages Security (ESORICS), volume 5789 of LNCS, pages 303–320.
295–304. ACM, 2015.
Springer, 2009.
[11] P. W. L. Fong and I. Siahaan.
control policies and their policy languages. In Proc. 16thACM Symposium on Access Control Models and Technolo-gies (SACMAT), pages 51–60. ACM, 2011.
[12] Y. Liu, K. P. Gummadi, B. Krishnamurthy, and A. Mislove.
Analyzing Facebook privacy settings: user expectations vs.
reality. In Proc. 2011 ACM SIGCOMM conference on In-ternet measurement conference (IMC), pages 61–70. ACM,2011.
[13] M. Mondal, Y. Liu, B. Viswanath, K. P. Gummadi, and A. Mislove. Understanding and specifying social access con-trol lists. In Proc. 10th Symposium on Usable Privacy andSecurity (SOUPS), pages 271–283. USENIX Association,2012.

Source: http://satoss.uni.lu/members/jun/papers/SACMAT16.pdf