[Paper Review] Trilateral Large-Scale OSN Account Linkability Study

Mishari Almishari, Ekin Oğuz+1 more|arXiv (Cornell University)|Oct 3, 2015

Authorship Attribution and Profiling16 references3 citations

TL;DR

This paper presents a trilateral large-scale stylometric study linking user accounts across three heterogeneous OSNs—Yelp (reviews), Twitter (microblogging), and Flickr (photo sharing)—using writing style analysis. Despite differing missions and content types, the authors achieve high linkability accuracy using a multi-level linkability framework (MLLF), demonstrating that cross-OSN privacy is significantly compromised even with only public textual data.

ABSTRACT

In the last decade, Online Social Networks (OSNs) have taken the world by storm. They range from superficial to professional, from focused to general-purpose, and, from free-form to highly structured. Numerous people have multiple accounts within the same OSN and even more people have an account on more than one OSN. Since all OSNs involve some amount of user input, often in written form, it is natural to consider whether multiple incarnations of the same person in various OSNs can be effectively correlated or linked. One intuitive means of linking accounts is by using stylometric analysis. This paper reports on (what we believe to be) the first trilateral large-scale stylometric OSN linkability study. Its outcome has important implications for OSN privacy. The study is trilateral since it involves three OSNs with very different missions: (1) Yelp, known primarily for its user-contributed reviews of various venues, e.g, dining and entertainment, (2) Twitter, popular for its pithy general-purpose micro-blogging style, and (3) Flickr, used exclusively for posting and labeling (describing) photographs. As our somewhat surprising results indicate, stylometric linkability of accounts across these heterogeneous OSNs is both viable and quite effective. The main take-away of this work is that, despite OSN heterogeneity, it is very challenging for one person to maintain privacy across multiple active accounts on different OSNs.

Motivation & Objective

• To investigate whether users can be reliably linked across three heterogeneous OSNs—Yelp, Twitter, and Flickr—despite differing content types and missions.
• To develop a scalable and accurate stylometric-based account linkability model that works with only publicly available textual data.
• To assess the privacy implications of maintaining multiple active accounts across different OSNs, especially under the assumption that users disable metadata and private messages.
• To evaluate the feasibility of extending linkability to larger account sets and to explore feature ordering and integration of additional stylometric features.

Proposed method

• The study employs a multi-level linkability framework (MLLF) that recursively narrows down candidate matches using stylometric features across OSNs.
• Stylometric features—including word frequency, sentence length, and function word usage—are extracted from user-generated text on Yelp, Twitter, and Flickr.
• The MLLF framework uses a hierarchical approach, applying different features at each level to progressively filter and rank potential matches.
• The method assumes no access to private data (e.g., geo-location, private messages), relying solely on publicly available text to establish a lower bound on adversary capability.
• Feature ordering is randomized across multiple runs to avoid bias from weak features being selected early, with results averaged over 10 runs.
• The framework is designed to scale linearly with the number of accounts, enabling application to large-scale datasets.

Experimental results

Research questions

• RQ1 Can user accounts be effectively linked across three heterogeneous OSNs—Yelp, Twitter, and Flickr—despite their differing content types and missions?
• RQ2 How accurate is stylometric analysis in linking accounts when only public textual data is available, without access to metadata or private messages?
• RQ3 Can the multi-level linkability framework (MLLF) achieve high scalability and accuracy across 100 to 100,000 user accounts?
• RQ4 How does feature ordering affect linkability performance, and can heuristics be developed to improve matching accuracy?
• RQ5 To what extent can additional features (e.g., hashtags, tags) or integration of profiles from similar OSNs (e.g., Flickr and Instagram) enhance linkability?

Key findings

• Stylometric analysis enables highly accurate cross-OSN account linkability across three heterogeneous platforms—Yelp, Twitter, and Flickr—despite differences in content type and mission.
• The multi-level linkability framework (MLLF) achieves high accuracy even when restricted to only publicly available textual data, indicating that privacy through compartmentalization across OSNs is ineffective.
• Linkability remains robust even without access to geo-location or private messages, suggesting that public text alone is sufficient for strong linkage attacks.
• The MLLF framework scales linearly with the number of accounts, with estimated execution time of ~2.5 minutes to link one unknown account to 1 million known accounts.
• Random feature ordering across 10 runs significantly improves reliability, and averaging results yields highly accurate linkability, suggesting robustness to feature selection bias.
• Future integration of additional stylometric features (e.g., from the Writeprints suite) and cross-OSN profile fusion (e.g., combining Twitter and Yelp profiles) may further enhance linkability performance.