Beyond Access Control: Managing Online Privacy via Exposure

Mainack Mondal, Peter Druschel, Krishna P. Gummadi, Alan Mislove2014 · DOI

TL;DR

This paper argues that traditional access control models are insufficient for managing online privacy, especially on social media. They propose a new model based on "exposure" – the set of people expected to eventually learn a piece of information – and suggest ways to build systems that allow users to better understand and control this exposure.

What Problem Does This Paper Address?

Online privacy is a growing concern, particularly with the rise of social media platforms. The dominant approach to managing privacy has been "access control," which dictates who can access information. However, the authors argue this model is fundamentally flawed for several reasons:

  • Difficulty of Specification: Users must precisely define who can and cannot access information beforehand, a task that's complex and prone to error.
  • Access vs. Actual Knowledge: Access control doesn't account for the ease or difficulty of finding information, nor does it distinguish between having permission to access something and actually doing so.
  • Redistribution and Inference: It fails to capture how information might be shared or how new information can be inferred from publicly available data.

The paper highlights real-world incidents (like Facebook's News Feed and Timeline, or Google Street View) where users felt their privacy was violated, even though no access control policies were technically broken. This suggests a need for a more nuanced understanding of privacy.

Methods

The core contribution of this paper is the proposal of a new model for online privacy: exposure.

  • Exposure Model: Instead of focusing on who can access information, exposure focuses on the set of principals (people) who are expected to eventually learn a piece of information. This is defined as the limit of the set of principals aware of the information over time, $E_I = \lim_{t \to \infty} P_I(t)$, where $P_I(t)$ is the set of principals aware of information $I$ at time $t$.
  • Factors Influencing Exposure: Exposure is influenced by two main factors:
  • The set of principals who meet the preconditions to learn the information (expertise, tools, initial leads).
  • The subset of those principals who are sufficiently motivated to actually learn it.
  • Exposure Control: The paper suggests moving from access control to "exposure control." This involves:
  • Predicting Exposure: Leveraging existing research on information propagation and dissemination to predict the likely exposure of content.
  • Making Predictions Transparent: System operators should make these predictions visible to users so they can calibrate their expectations.
  • Controlling Exposure: Providing users with mechanisms to tune their exposure, such as opting in/out of dissemination channels or adjusting access controls to influence the predicted viewer set.
  • Tripwires: Implementing mechanisms that automatically make information inaccessible if its actual exposure deviates significantly from predictions, notifying the user.

Datasets

The paper describes a small-scale study using data from three real-world scenarios to evaluate the accuracy of exposure predictions:

  • Scenario 1: Public YouTube video views.
  • Scenario 2: Views for posts in Facebook groups.
  • Scenario 3: "Likes" on personal Facebook photos shared with friends.

The authors collected data on daily views/likes for a set of videos/posts/photos and used historical data to predict future popularity.

Key Insights and Ideas

  • Access Control is a Poor Proxy for Privacy: The paper strongly argues that focusing solely on who can access information misses crucial aspects of user privacy, particularly in dynamic online environments.
  • Exposure is More Intuitive: The concept of "exposure" – who is likely to actually learn something – is a more direct and intuitive measure of privacy implications than access control.
  • Surprise is a Key Driver of Privacy Violations: Users feel their privacy is violated when the actual exposure of information differs significantly from their expectations, especially if unexpected people learn it.
  • Social Media Amplifies Exposure Issues: Features like Facebook's News Feed and Timeline, while not breaking access rules, dramatically increase exposure by making information more readily discoverable or "pushed" to users.
  • Data Aggregation Creates New Privacy Risks: Linking publicly available pieces of information can create new, inferred information with higher exposure than the individual components.
  • Predicting Exposure is Feasible and Useful: The paper demonstrates that existing data and models can be used to predict content popularity and, by extension, exposure, with reasonable accuracy.
  • User Control is Essential: Simply predicting exposure isn't enough; users need tools to actively manage and control it to align with their expectations.

Key Results

The study on exposure prediction showed promising results:

  • Scenario 1 (Public YouTube views): For 75% of videos, the relative error in prediction was less than 0.1 (meaning the actual value was within ±10% of the predicted value).
  • Relative Error Decreases with Restricted Access: The accuracy of predictions (measured by relative error) improved as access controls became more restrictive, moving from public YouTube videos (highest error) to Facebook group posts, and finally to personal Facebook photos shared with friends (lowest error). This suggests that tighter access controls make exposure more predictable.
  • Niche Content is More Predictable: The paper observes that the popularity of less popular (niche) content stabilizes and becomes more predictable over time compared to highly popular content.

Limitations and Future Work

  • Model Limitations: The paper acknowledges that no model can perfectly predict occasional sharp, disruptive jumps in popularity (e.g., content going viral).
  • Implementation Challenges: While the paper proposes mechanisms for exposure control, a concrete implementation and deployment-based study is needed to fully explore the research questions.
  • Future Work: The authors suggest further research into:
  • Concrete implementations of exposure control mechanisms.
  • Deployment-based studies to gather empirical data and refine models.
  • Developing more sophisticated models for predicting information propagation.
  • Investigating the overheads associated with fine-tuning exposure.
  • Exploring how users can organize information into groups for easier exposure management.

Summary updated

Abstract

We posit that access control, the dominant model for modeling and managing privacy in today’s online world, is fundamentally inadequate. First, with access control, users must a priori specify precisely who can or cannot access information by enumerating users, groups, or roles—a task that is difficult to get right. Second, access control fails to separate who can access information from who actually does, because it ignores the difficulty of finding information. Third, access control does not capture if and how a person who has access to some information redistributes that information. Fourth, access control fails to account for information that can be inferred from other, public information. We present exposure as an alternate model for information privacy; exposure captures the set of people expected to learn an item of information eventually. We believe the model takes an important step towards enabling users to model and control their privacy effectively.
BibTeX
@misc{mondal2014beyond,
  author = {Mainack Mondal and Peter Druschel and Krishna P. Gummadi and Alan Mislove},
  title = {{Beyond Access Control: Managing Online Privacy via Exposure}},
  year = {2014},
  doi = {10.14722/usec.2014.23046},
}
Figure