Skip to main content
eScholarship
Open Access Publications from the University of California

Combating Attacks and Abuse in Large Online Communities

  • Author(s): Wang, Gang
  • Advisor(s): Zhao, Ben Y
  • Zheng, Haitao
  • et al.
Abstract

Internet users today are connected more widely and ubiquitously than ever before. As a result, various online communities are formed, ranging from online social networks (Facebook, Twitter), to mobile communities (Foursquare, Waze), to content/interests based networks (Wikipedia, Yelp, Quora). While users are benefiting from the ease of access to information and social interactions, there is a growing concern for users' security and privacy against various attacks such as spam, phishing, malware infection and identity theft.

Combating attacks and abuse in online communities is challenging. First, today’s online communities are increasingly dependent on users and user-generated content. Securing online systems demands a deep understanding of the complex and often unpredictable human behaviors. Second, online communities can easily have millions or even billions of users, which requires the corresponding security mechanisms to be highly scalable. Finally, cybercriminals are constantly evolving to launch new types of attacks. This further demands high robustness of security defenses.

In this thesis, we take concrete steps towards measuring, understanding, and defending against attacks and abuse in online communities. We begin with a series of empirical measurements to understand user behaviors in different online services and the unique

security and privacy challenges that users are facing with. This effort covers a broad set of popular online services including social networks for question and answering (Quora), anonymous social networks (Whisper), and crowdsourced mobile communities (Waze). Despite the differences of specific online communities, our study provides a first look at their user activity patterns based on empirical data, and reveals the need for reliable mechanisms to curate user content, protect privacy, and defend against emerging attacks.

Next, we turn our attention to attacks targeting online communities, with focus on spam campaigns. While traditional spam is mostly generated by automated software, attackers today start to introduce "human intelligence" to implement attacks. This is malicious

crowdsourcing (or crowdturfing) where a large group of real-users are organized to carry out malicious campaigns, such as writing fake reviews or spreading rumors on social media. Using collective human efforts, attackers can easily bypass many existing defenses (e.g.,

CAPTCHA). To understand the ecosystem of crowdturfing, we first use measurements to examine their detailed campaign organization, workers and revenue. Based on insights from empirical data, we develop effective machine learning classifiers to detect crowdturfing

activities. In the meantime, considering the adversarial nature of crowdturfing, we also build practical adversarial models to simulate how attackers can evade or disrupt machine learning based defenses.

To aid in this effort, we next explore using user behavior models to detect a wider range of attacks. Instead of making assumptions about attacker behavior, our idea is to model normal user behaviors and capture (malicious) behaviors that are deviated from norm. In this way, we can detect previously unknown attacks. Our behavior model is based on detailed clickstream data, which are sequences of click events generated by users when using the service. We build a similarity graph where each user is a node and the edges are weighted

by clickstream similarity. By partitioning this graph, we obtain "clusters" of users with similar behaviors. We then use a small set of known good users to "color" these clusters to differentiate the malicious ones. This technique has been adopted by real-world social networks (Renren and LinkedIn), and already detected unexpected attacks. Finally, we extend clickstream model to understanding more-grained behaviors of attackers (and real users), and tracking how user behavior changes over time.

In summary, this thesis illustrates a data-driven approach to understanding and defending against attacks and abuse in online communities. Our measurements have revealed new insights about how attackers are evolving to bypass existing security defenses today. In

addition, our data-driven systems provide new solutions for online services to gain a deep understanding of their users, and defend them from emerging attacks and abuse.

Main Content
Current View