To Join or Not to Join The Illusion

To Join or Not to Join: The Illusion of Privacy in Social Networks with Mixed Public and Private User Profiles By Elena Zheleva, Lise Getoor Presented by Ionut Trestian

Privacy is important ! (1)

Privacy is important ! (2)

Privacy is important ! (3) • Sometimes it’s not the user who accidently exposes private information • Groups, organizations that the users belong to might expose information accidentally or not

Contributions • Identify novel social network attacks • We show that such attacks can be carried out even with limited information • We evaluate our attacks on real social network data (Flickr, Facebook, Dogster and Bib. Sonomy) • Discuss how our study affects anonymization of social networks

Types of attacks • Attacks without links and groups {BASIC} – Pick the most probable attribute from public profiles – Simple, use as a baseline • Privacy attacks using links • Privacy attacks using groups • Privacy attacks using links and groups

Privacy attacks using links • Friend-aggregate model (AGG) – Pick the most probable attribute value from friends • Collective classification model (CC) – Iterative classification • Flat-link model (LINK) – Traditional classifiers, Bayes etc • Blockmodeling attack (BLOCK) – Obtain blocks (clusters of users) and find where the user belongs

Privacy attacks using groups • Groupmate-link model (CLIQUE) – Assume group members are friends • Group-based classification model (GROUP) – Consider groups as features – Not all groups are relevant

Privacy attacks using links and groups • Combine flat-link and group-based classification models into one • LINK-GROUP • Can use any traditional classifier

Experiments - Data • Flickr - 9, 179 users from 55 countries (47, 754 groups) • Facebook – 1, 598 users – political views • Dogster – 2, 632 dogs – 1, 042 groups • Bib. Sonomy – 31, 175 users + tags

Results (1) • 50% private profiles

Results GROUP (2)

Results GROUP (3)

Results GROUP (4)

Results GROUP (5)

Results GROUP (6)

Results GROUP (7)

Discussion • Joining heterogeneous groups preserves privacy better • Display Group information only to friends • Remove homogeneous groups

Thank you ! Questions ?