Understanding and Specifying Social Access Control Lists

Mainack Mondal, Yabing Liu, Bimal Viswanath, Krishna P. Gummadi, Alan MisloveSOUPS 2014 · OpenReview

TL;DR

This paper presents the first large-scale study of how users actually specify privacy settings (Social Access Control Lists or SACLs) on Facebook. They found that SACLs are surprisingly common, but their membership is hard to predict from user attributes or network structure, suggesting current automatic methods are insufficient. The study also highlights the significant user effort involved in creating SACLs and proposes simple, history-based methods to substantially reduce this burden.

What Problem Does This Paper Address?

Online Social Networks (OSNs) like Facebook allow users to share vast amounts of personal content. While default privacy settings exist, users often need to restrict access to sensitive content to specific groups of friends. This is done using Social Access Control Lists (SACLs), which are essentially custom lists of friends who are allowed to see a piece of content. However, it's unclear how users actually create and use these SACLs. Designing effective privacy management tools requires understanding this "ground truth" of user behavior, rather than relying on hypothetical scenarios or small-scale surveys. The authors aim to fill this gap by studying real-world SACL usage on Facebook.

Methods

The researchers developed a Facebook application called "Friendlist Manager" (FLM) to collect data on user-specified SACLs. This app requested user consent to access their privacy settings, friend lists, and content sharing information. To analyze SACL membership, they employed several techniques:

  • Attribute-based clustering: Grouping friends based on shared profile information (workplace, education, etc.).
  • Network community detection: Identifying groups of friends who are densely connected to each other.
  • Activity-based clustering: Grouping friends based on their interaction frequency with the user.

They then compared these automatically generated clusters to the users' actual SACLs using metrics like F-score (for match accuracy), normalized entropy (to measure distribution across clusters), and Adjusted Rand Index (ARI) (to compare against random groupings). To assess user effort, they measured the "overhead" of specifying SACLs, defined as the number of terms (individual friends or lists) used. They also explored methods to reduce this overhead, such as using the "last-used setting" and leveraging a user's "past history" of frequently used SACLs.

Datasets

The study utilized a dataset collected from 1,165 users of a Facebook application ("Friendlist Manager"). This dataset comprised 212,753 pieces of uploaded content for which users had specified SACLs. The data included user profile attributes, friend lists, and the specific SACLs used for each piece of content. This dataset represents a novel "ground truth" of real-world SACL usage, distinct from data derived from surveys or hypothetical user studies.

Key Insights and Ideas

  • SACLs are common and important: Contrary to what might be assumed, a significant portion of content (17.6%) is shared using SACLs, indicating they play a crucial role in users' privacy management for sensitive information.
  • Membership is surprisingly unpredictable: SACL members show little correlation with profile attributes or network structure. This is a key finding that challenges existing assumptions and the effectiveness of current automatic SACL generation methods.
  • User effort is high, but reducible: Specifying SACLs can be cognitively demanding. However, simple interventions like remembering the "last-used setting" or suggesting frequently used SACLs from a user's history can dramatically reduce this burden.
  • "Ground truth" matters: The study's findings differ from previous work because it uses actual SACL specifications, not user-reported preferences. This highlights the importance of studying real-world behavior for designing effective privacy tools.

Key Results

  • Usage: 67.8% of users in the study shared at least one piece of content using SACLs.
  • Content Distribution: SACLs were used across all content types, with 10-20% of most content types being shared with SACLs.
  • Membership Predictability: Only about 15% of attribute-based clusters, 21% of network communities, and 47% of activity-based clusters showed a strong match (F-score > 0.6) with SACLs. ARI analysis also showed limited correlation with random groupings.
  • SACL Size: The distribution of SACL sizes varied, with a significant portion being "include few" (0-5% of friends) or "exclude few" (95-100% of friends), but a plurality were "include subset" (5-95% of friends).
  • User Overhead Reduction:
    • Using the "last-used setting" reduced user overhead by over 50% for nearly half of users.
  • Leveraging a user's "past history" (top 5 most used SACLs) could reduce overhead for 86% of users, covering over 80% of their content.

Limitations and Future Work

  • User Bias: The study acknowledges that users who installed the FLM app might be more privacy-aware than the general Facebook population. Future work could explore how to mitigate this bias.
  • Understanding Motivation: The study describes how users create SACLs but not why. Future research could investigate user motivations, such as dissatisfaction with defaults or the desire to control audiences for specific content.
  • SACL Accuracy: The paper doesn't definitively know if the SACLs users create perfectly reflect their ideal privacy preferences. Users might misunderstand settings or face cognitive burdens that lead to suboptimal choices.
  • Overhead Measurement: The definition of overhead (number of terms) is a simplification. Future work could explore other dimensions of cognitive overhead through interviews.
  • Moving Target: Facebook's privacy interface changes frequently. The authors plan to repeat their analysis to capture how these changes affect user behavior.

Summary updated

BibTeX
@inproceedings{mondal2014understanding,
  author = {Mainack Mondal and Yabing Liu and Bimal Viswanath and Krishna P. Gummadi and Alan Mislove},
  title = {{Understanding and Specifying Social Access Control Lists}},
  booktitle = {SOUPS},
  year = {2014},
  url = {https://openreview.net/forum?id=87oFkYeKiu},
}
Figure