Thwarting fake accounts by predicting their victims

Yazan Boshmaf ; Dionysios Logothetis ; Georgos Siganos ; Matei Ripeanu ; Konstantin Beznosov

24 March 2014

Abstract: Traditional fake account detection systems employed by today's online social networks rely on either features extracted from user activities, or ranks computed from the underlying social graph. We herein present a system that integrates both approaches to deliver a more resilient defense mechanism that is both efficient and effective. We present a two-phase, iterative technique to achieve this integration. First, we leverage the insight that harmful fake accounts connect with other users (i.e., victims) before mounting subsequent attacks. We therefore train a classifier to predict these victims using features extracted from the activities of known, non-fake accounts. Second, we observe that actual victims are located at the borderline between two subgraphs, effectively separating harmful fake accounts from other accounts in the social graph. We take advantage of this observation by using the predicted victims as "deflection points" for a short random walk that starts from a known, non-fake account that is not a victim. By ranking accounts based on their landing probability, we guarantee that most of the fake accounts have a strictly lower rank than non-fake accounts. The results of our experiments show that our technique can help in reducing the number of victims while providing a more robust ranking for fake accounts detection.

Keyword(s): Social network security ; Online social networks ; Fake accounts detection ; Graph theory ; Machine learning

Published in: Boshmaf et al. Thwarting fake accounts by predicting their victims. Invited talk at AAAI 2014 Spring Symposia, Social Hacking and Cognitive Security on the Internet and New Media, Stanford, CA, March, 2014.:

The record appears in these collections:
Network Security

 Record created 2014-04-01, last modified 2014-06-02

Download fulltextPDF Download fulltextPDF (PDFA)
Rate this document:

Rate this document:
(Not yet reviewed)