Towards robust trust establishment in web-based social networks with socialtrust

James Caverlee, Ling Liu, Steve Webb

Abstract

Social networks face significant threats from malicious actors who exploit the "trust gap" inherent in large-scale digital interactions. This paper presents SocialTrust, a reputation-based framework designed to mitigate these risks. By decoupling relationship quality from user trustworthiness and implementing a dynamic feedback loop based on control theory, SocialTrust provides a resilient mechanism for identifying reliable information. Experimental results on a large-scale MySpace dataset demonstrate that SocialTrust maintains high precision even when up to 70% of network participants are malicious.

Introduction

In the mid-2000s, the rapid expansion of social networks like MySpace and Facebook created a significant "trust gap." While these platforms facilitated global connectivity, they remained vulnerable to phishing, spam, and malware. The authors identify three structural weaknesses: malicious infiltration facilitated by low barriers to entry, the "small world" nature of networks that places threats in close proximity to targets, and the limited network view of individual users. Traditional trust models, such as PageRank, often fail in these environments because they cannot distinguish between popularity and trustworthiness, allowing malicious actors to manipulate scores through dense sub-networks of fake accounts.

The SocialTrust Framework

The core of the SocialTrust model is the treatment of trust as a dynamic signal rather than a static value. It introduces two primary innovations:

1. PID-Controller Approach

Inspired by control theory, SocialTrust defines a user's score as a weighted sum of three components:

  • Proportional: Current behavioral quality.
  • Integral: Historical behavior, preventing "exit scams" where users turn malicious after building a positive reputation.
  • Derivative: The rate of change in behavior, allowing the system to quickly penalize sudden malicious spikes.

2. Recursive Trust Metric ($Tr_q$)

SocialTrust utilizes a recursive formula that balances relationship quality ($L$) with personalized feedback ($F$). Unlike PageRank, which treats all links equally, SocialTrust evaluates the quality of links pointing to a user and incorporates actual interaction feedback to update trust scores dynamically.

Experimental Evaluation

The researchers evaluated SocialTrust using a directed graph harvested from MySpace, consisting of approximately 5.2 million nodes and 19.1 million relationship links. The simulation modeled a scenario where users search for information, contrasting "legitimate" users with "malicious" participants who provide harmful or irrelevant responses.

Key Results

SocialTrust demonstrated superior resilience compared to industry standards like PageRank and TrustRank.

Comparison of SocialTrust against PageRank and TrustRank under varying levels of malicious participation.

As illustrated in the figure above, in a network where 70% of participants were malicious, SocialTrust maintained a relative precision of nearly 0.8. In contrast, the precision of PageRank and TrustRank dropped to roughly 0.3 under the same conditions. This resilience is attributed to SocialTrust's "closed-loop" system, which uses interaction results to update graph weights dynamically.

Conclusion and Future Work

SocialTrust provides a robust framework for maintaining information quality in adversarial social networks. By separating relationship quality from trust and employing feedback loops, it effectively mitigates the impact of sybil attacks and malicious clusters. Future research will explore context-aware trust models—allowing for different trust levels across various domains—and the incorporation of more sophisticated relationship features, such as relationship duration and value.

Summary updated

Abstract

We propose the SocialTrust framework for tamper-resilient trust establishment in online social networks. Two of the salient features of SocialTrust are its dynamic revision of trust by (i) distinguishing relationship quality from trust; and (ii) incorporating a personalized feedback mechanism for adapting as the social network evolves.
BibTeX
@inproceedings{caverlee2008towards,
  author = {James Caverlee and Ling Liu and Steve Webb},
  title = {{Towards robust trust establishment in web-based social networks with socialtrust}},
  booktitle = {The Web Conference},
  year = {2008},
  doi = {10.1145/1367497.1367707},
}
Figure