PaperlayerView exact revision on arXiv

Source-linked AI summary

Friend or Foe? Fake Profile Identification in Online Social Networks

Michael Fire, Dima Kagan, Aviad Elyashar, Yuval Elovici

arXiv:1303.3751v1cs.SIphysics.soc-ph

TL;DR

Facebook users may expose personal information through their connections, applications, and privacy settings, while fake profiles can threaten their security and privacy. The paper presents SPP, combining friend analysis, privacy-setting assistance, application alerts, and machine-learning classifiers. Rotation-Forest classifiers achieved an AUC of 0.948 and a false-positive rate of 15.8% on the Fake profiles dataset.

  • Problem

    Facebook users may disclose personal information through friend connections, applications, and privacy settings, while attackers may use fake profiles to conceal activity and threaten users.

    Source: §1 · pp. 1–2

  • Method

    SPP combines friend analysis for restricting suspected fake profiles, one-click privacy-setting options, application-count alerts, and machine-learning classifiers developed from software-collected data.

    Source: §1 · p. 2

  • Results

    AUC 0.948 and false-positive rate 15.8% on the Fake profiles dataset were achieved by the best-performing Rotation-Forest classifiers; in the general link-restriction scenario, one classifier reached 91% precision at 100 and 94% at 500.

    Source: §6 · pp. 21–22

  • Takeaways & Limitations

    The reported classifier precision supports their potential use by social-network users and administrators to identify fake profiles among network profiles.

    Source: §6 · p. 22

    Unexpected downloads and usage exceeded server capacity, and the initial version did not support all browsers, preventing some installers from using SPP on demand.

    Source: §1 · p. 2

Abstract

from arXiv · show

The amount of personal information unwillingly exposed by users on online social networks is staggering, as shown in recent research. Moreover, recent reports indicate that these networks are infested with tens of millions of fake users profiles, which may jeopardize the users' security and privacy. To identify fake users in such networks and to improve users' security and privacy, we developed the Social Privacy Protector software for Facebook. This software contains three protection layers, which improve user privacy by implementing different methods. The software first identifies a user's friends who might pose a threat and then restricts this "friend's" exposure to the user's personal information. The second layer is an expansion of Facebook's basic privacy settings based on different types of social network usage profiles. The third layer alerts users about the number of installed applications on their Facebook profile, which have access to their private information. An initial version of the Social Privacy Protection software received high media coverage, and more than 3,000 users from more than twenty countries have installed the software, out of which 527 used the software to restrict more than nine thousand friends. In addition, we estimate that more than a hundred users accepted the software's recommendations and removed at least 1,792 Facebook applications from their profiles. By analyzing the unique dataset obtained by the software in combination with machine learning techniques, we developed classifiers, which are able to predict which Facebook profiles have high probabilities of being fake and therefore, threaten the user's well-being. Moreover, in this study, we present statistics on users' privacy settings and statistics of the number of applications installed on Facebook profiles...

1 Introduction

Facebook users can expose personal information through friend connections, applications, and privacy settings, while fake profiles may use that information to threaten users. The paper introduces SPP to help protect Facebook users and reports deployment, classifier, and privacy-statistics results.

Motivation

  • Users may disclose personal details, and friend requests from strangers or third-party applications can expose that information; many users' privacy settings do not match their security expectations.The cited study found that an average of 80% of studied Facebook users accepted requests from unknown people when they shared more than 11 mutual friends.Source: §1 · pp. 1–2
  • Fake profiles can help attackers conceal their activity; Facebook estimated 8.7% of accounts (83.09 million) were not real profiles and 1.5% (14.32 million) were undesirable accounts.Exposed information could be used for customized spam or blackmail.Source: §1 · p. 2

Contribution

  • SPP combines one-click privacy-setting controls, warnings about installed applications, and a Facebook application that identifies suspected fake-profile friends for restriction.Users can restrict access without removing those friends from their list.Source: §1 · p. 2

Deployment

  • 3,017 users from more than 20 countries installed SPP; 527 used it to restrict 9,005 friends in less than four months.At least 111 users were estimated to have removed more than 1,792 applications following the add-on recommendation; 67 users' privacy settings were collected.Source: §1 · p. 2

Results

  • AUC 0.948 and false-positive rate 15.8% on the Fake profiles dataset were achieved by the best-performing Rotation-Forest classifiers.The classifiers were developed using data collected by SPP and machine learning techniques.Source: §1 · p. 2

Deployment

  • SPP's initial deployment had server-capacity and browser-compatibility constraints, so some installers could not use the software on demand.These limits followed unexpected high downloads and usage.Source: §1 · p. 2

Results

  • 10.68% of 1,676 examined users had more than 100 Facebook applications installed, and 30.31% had at least 40.Among 67 collected privacy settings, most personal information was exposed to friends, leaving it exposed to fake friends.Source: §1 · p. 3

2 Related Work

Prior work addressed social-network security through platform features, commercial tools, and research on threats such as spam, malicious applications, and fake profiles. This paper combines user-side privacy protection with machine-learning classifiers built from data collected through SPP.

2.1 Online Social Network Security and Privacy

  • SPP offers Facebook privacy-protection methods, while data collected by the software and machine learning support classifiers for identifying fake profiles.This contribution bridges user-facing protection and profile identification.Source: §2.1 · pp. 3–4

2.2 Online Social Networks and Machine Learning

  • Earlier work applied machine learning to link prediction, spammer detection, event detection, user-property prediction, and fake-user identification.The cited studies include work on social-network security and user-property prediction.Source: §2.2 · p. 4
  • Unlike studies using broad network topology, this work predicts fake profiles from real-time user-view data, including connection and meta-content features, to analyze friends lists with fewer resources and without invading friends' privacy.The task is distinguished from predicting link strength or signed links [24].Source: §2.2 · p. 4

3 Social Privacy Protector Architecture

SPP links a Facebook application, a Firefox add-on, and an HTTP server to provide friend analysis, privacy-setting assistance, and application-count warnings. Its components rank friends for review, offer profile-based privacy options, and process results through the server.

3 Social Privacy Protector Architecture

  • The architecture combines a Friends Analyzer Facebook application, an SPP Firefox Add-on, and an HTTP server that analyzes, stores, and caches results.The three components work together to provide privacy protection.Source: §3 · p. 4, Figure 1 · p. 5

3.1 The Friends Analyzer Facebook Application

  • The Friends Analyzer scans a user's friends list and assigns credibility scores using heuristics or machine-learning algorithms based on connection strength.It sorts friends by score, placing the lowest-scoring friends—considered most likely fake—at the top for possible restriction.Source: §3.1 · pp. 4–5

3.2 Social Privacy Protector Firefox Add-on

  • The Firefox add-on warns users about installed applications and surfaces two Friends Analyzer results; it also offers profile-type privacy settings that can be changed with one click.The three options are Celebrity, Recommended, and Kids settings.Source: §3.2 · pp. 5–6, Figure 2 · p. 6

3.3 HTTP Server

  • The HTTP server connects the add-on and Facebook application, analyzes friends lists, and caches results; it stores only a minimal number of features encrypted with RC4.Source: §3.3 · p. 6

4 Methods and Experiments

The methods and experiments describe deploying the Social Privacy Protector to collect privacy data and using the resulting links and features to develop and evaluate fake-profile classifiers. Evaluation includes a connection-strength heuristic, supervised learning across link datasets, and two precision@k tests of the highest-AUC classifier.

4.1 Deploying Social Privacy Protector - Initial Version

  • The initial SPP deployment collected link data to develop and assess the heuristic, while also recording users’ privacy settings and installed-application counts.Collected settings covered default privacy, lookup, address sharing, messaging, friend requests, tag suggestions, and birthday visibility; the study used these data to examine privacy and exposure risks.Source: §4 · p. 6, §4.1 · p. 10
  • SPP ranked friends by a connection-strength heuristic and recommended the 10% with the lowest scores for restriction.The score adds chat messages, twice the shared groups, posts, tagged photos and videos, and 1,000 times the family indicator; mutual likes and wall comments were excluded because they were too slow to compute in real time.Source: §4.1 · p. 9

4.2.1 Datasets and features

  • Supervised learning used three link datasets to identify likely fake profiles, predict user-selected restrictions, or replace the heuristic with recommendations covering all links.The fake-profiles dataset combines recommended-but-unrestricted and all unrestricted links; the friends-restriction dataset combines alphabetically restricted and all unrestricted links, while the all-links dataset contains all four link sets.Source: §4.2.1 · p. 11
  • The datasets represented different classification goals, and each link received 15 features, including a Jaccard coefficient measuring shared friends relative to distinct friends.Other features included connection signals and ratios, friend counts, and a private-profile indicator inferred when a friend count was unavailable despite a positive mutual-friend count.Source: §4.2.1 · p. 12, §4.2.2 · p. 13

4.2.2 Classifiers Construction and Evaluation

  • Classifier construction balanced restricted and unrestricted links by undersampling, compared seven WEKA algorithms against OneR, and used 10-fold cross-validation with AUC, F-measure, and true- and false-positive rates.Feature importance was also assessed with WEKA information gain; classifier types were matched to the fake-profile, friends-restriction, and all-links datasets.Source: §4.2.2 · p. 13
  • The first evaluation trained on two-thirds of each balanced dataset and ranked held-out profiles by predicted restriction probability to measure precision@k.The classifier selected was the one with the highest AUC in the preceding evaluations.Source: §4.2.2 · p. 13
  • The second evaluation trained on links from all other SPP users, tested on one selected user’s links, and averaged precision at different k values across users.The training data was balanced using the previously described undersampling method.Source: §4.2.2 · pp. 13–14
  • Together, the methods assessed recommendation precision from both SPP users’ and social-network administrators’ perspectives.Source: §4.2.2 · p. 14

5 Results

The initial SPP deployment attracted thousands of users and yielded three imbalanced link datasets; classifiers and add-on logs characterized restriction recommendations and users’ privacy-related behavior.

5.1 Collected Datasets

  • The Fake-profiles, Friends-restriction, and All links datasets contained 141,146, 144,431, and 151,825 links, with restricted-link shares of 2.03%, 4.25%, and 6.01%.All three datasets were overwhelmingly imbalanced.Source: §5.1 · p. 14, Table 1 · p. 14
  • The Connection-Strength heuristic averaged 33.6% precision at 1 and 11% at 100; only 10% of friends with a score of 3 were restricted.Restriction rates declined sharply as Connection-Strength increased.Source: §5.1 · p. 15, Figure 5 · p. 16

5.2 Classifiers’ Results

  • Rotation-Forest achieved the highest AUC on all three datasets and considerably outperformed the OneR baseline.Source: §5.2 · p. 16, Figure 8 · p. 17, Table 3 · p. 18
  • At precision@100, Rotation-Forest reached 98% on the Friends restriction dataset, 91% on the Fake profiles dataset, and 91% on the All links dataset.The classifier’s precision@500 was 94% on Fake profiles, 91% on Friends restriction, and 88% on All links.Source: §5.2 · p. 16, Figure 9 · p. 19

5.3 Add-on Results

  • At initial add-on installation, 10.68% of users had more than 100 Facebook applications installed, while 34.96% had fewer than ten.Source: §5.3 · p. 19, Figure 11 · p. 20
  • Among 626 users observed a day after add-on installation, 111 (17.73%) removed 1,792 applications; across the full period, 335 users removed 5,537.On average, the 111 users removed 34.7% of their installed applications.Source: §5.3 · pp. 19–20
  • Among 67 users with collected privacy settings, 74.62% set defaults to expose information to everyone, and 14 (20.9%) changed their settings during the study period.Source: §5.3 · p. 20, Table 5 · p. 21

6 Discussion

Rotation-Forest classifiers performed best across the tested datasets, with strong results for fake-profile classification and high precision in the general link-restriction scenario. The discussion also reports limits in the restricted-friends dataset, persistent privacy exposure, and user activity around Facebook applications.

6 Discussion

  • AUC of 0.948 and a false-positive rate of 15.8% on the Fake profiles dataset were achieved by the best-performing Rotation-Forest classifiers.They performed best among the tested machine-learning algorithms on all datasets.Source: §6 · p. 21
  • At 1, Rotation-Forest classifiers reached 21% and 20% average users precision on All links and Fake-profiles datasets versus 34% for Connection-Strength; at 20, results were 21%, 20%, and 22%, respectively.Rotation Forest provides a generic restriction-recommendation model and per-link probabilities without comparing links within each user; the heuristic’s 31.1% top-ranked precision did not provide a generic method with high true-positive and low false-positive rates.Source: §6 · pp. 21–22
  • 14% average users precision at 1 on the Restricted friends dataset exceeded the 4.25% random-guessing baseline, but the authors attribute low performance to real friends and reasons not captured by extracted features.The authors describe that explanation as an assumption.Source: §6 · p. 22
  • 91% precision at 100 and 94% at 500 were reported for the Fake profiles classifier, while the Friends restriction classifier reached 98% precision at 100.These results concern the general scenario of predicting which links to restrict among all users’ links.Source: §6 · p. 22
  • Common-friends and Jaccard’s-Coefficent were the most useful features across datasets, while Is Family and Tagged-Video-Number were less useful.The authors suggest dropping the latter features could improve run time without significantly affecting results.Source: §6 · p. 22
  • 30.31% of 1,676 examined users had at least 40 applications and 10.68% had more than 100; 28.2% of 111 add-on users removed over half their applications.Separately, 389 of the 1,676 users added applications, averaging 1.91 new installations per week.Source: Figure 13 · p. 22, §6 · p. 23
  • Almost all examined users made their information available to friends, and most set default privacy to everyone, leaving information publicly viewable..Source: §6 · p. 23
  • SPP assisted users with restricting friends and removing unwanted applications, but generally did not improve their privacy settings.Source: §6 · p. 23

7 Conclusions

The paper presents SPP’s three privacy-protection layers and evaluates friend-restriction recommendations, fake-profile classifiers, and Facebook privacy and application patterns. Classifiers performed best on AUC and false-positive rates across three datasets, while their false-positive rates remained relatively high.

7 Conclusions

  • SPP combines friend restriction recommendations, warnings about installed applications, and one-click privacy-setting adjustments.The friend-recommendation function may use a heuristic or a machine-learning classifier.Source: §7 · p. 23
  • 527 of 3,017 users restricted 9,005 friends, but the Connection-Strength heuristic did not provide a general method for deciding which links to restrict.Users restricted 30.87% of friends shown first in the Restriction interface.Source: §7 · p. 23
  • Rotation-Forest had the best AUC and false-positive rates on all three datasets, while its classifiers showed good average users precision at k but relatively high false-positive rates.The classifiers were also presented as a method that could assist Facebook administrators in identifying fake profiles across the network.Source: §7 · pp. 23–24
  • 30.31% of 1,676 users had at least forty installed applications, and 111 add-on users removed at least 1,792 applications.Users installed an average of 1.91 new applications per week; statistics from 67 users also showed that many exposed private information to friends or the public.Source: §7 · p. 24
  • Future SPP work was planned to add browser support, remove less useful features, and grow the classifier training set by helping users restrict more fake profiles.The proposed browsers included Chrome and Internet Explorer; the Tagged-Video-Number feature was named for removal.Source: §7 · p. 24
  • Proposed research directions include richer topological features, oversampling instead of undersampling, testing on other networks, and studying SPP’s educational role and user awareness.The paper also proposes deeper analysis of the collected datasets and testing the algorithms on other online social networks.Source: §7 · p. 24
Loading 1303.3751v1…