Follow My Recommendations: A Personalized Privacy Assistant for Mobile App Permissions

Bin Liu, Mads Schaarup Andersen, Florian Schaub, Hazim Almuhimedi, Shikun Zhang, Norman M. Sadeh, Yuvraj Agarwal, Alessandro AcquistiSOUPS 2016 · OpenReview

TL;DR

This paper introduces a Personalized Privacy Assistant (PPA) that helps users manage mobile app permissions by learning their preferences and offering tailored recommendations. The PPA significantly increases user adoption of privacy-conscious settings, demonstrating a practical approach to improving mobile privacy management.

What Problem Does This Paper Address?

Modern smartphones offer millions of apps, each requesting various permissions to access sensitive user data and device functionalities. While platforms provide permission controls, the sheer volume of decisions overwhelms users, leading to a lack of awareness and discomfort with their privacy settings. Existing solutions often lack personalization or operationalization, leaving a gap in effectively guiding users toward their desired privacy configurations. This paper tackles the challenge of making mobile app permission management more manageable and aligned with individual user preferences.

Methods

The core of this work is the development of a Personalized Privacy Assistant (PPA) that operates in two main phases:

  1. Profile Building:
    • Enhanced Permission Manager: A modified Android permission manager was used to collect real-world permission settings, access frequency, and inferred purpose of access (using PrivacyGrade).
  2. Clustering: Hierarchical clustering was applied to user feature vectors (derived from app categories, permissions, and purposes) to identify distinct groups of users with similar privacy preferences, forming "privacy profiles."
  3. PPA Operation:
    • Interactive Profile Assignment: For new users, the PPA uses a dynamically generated decision tree to ask a small number of tailored questions (up to 5) about their privacy preferences, contextualized by their installed apps. This process assigns the user to one of the learned privacy profiles.
  4. Personalized Recommendations: Based on the assigned profile and the user's installed apps, a scalable SVM classifier generates specific recommendations for permission settings (primarily to deny access).
  5. Privacy Nudges: Daily "nudges" were used to remind users about app permission usage and encourage them to review and adjust their settings, particularly in the evaluation phase.

Datasets

  • Profile Building Dataset: Collected from 84 Android users with rooted devices using the enhanced permission manager. This dataset included permission settings, app categories, and inferred purposes.
  • Evaluation Dataset: Consisted of 72 participants (49 in the treatment group, 23 in the control group) who used the PPA in a field study.
  • External Dataset (for downsampling comparison): Lin et al.'s dataset [28] of 21,657 records from 725 MTurkers was used to validate the stability of learned profiles with smaller sample sizes.

Key Insights and Ideas

  • Privacy Profiles are Feasible and Useful: The study demonstrates that it's possible to learn distinct privacy profiles from a relatively small set of users and that these profiles can effectively guide personalized recommendations.
  • Contextualization is Key: Tailoring questions for profile assignment based on a user's installed apps significantly improves the relevance and usability of the process.
  • Recommendations Drive Adoption: Users are highly likely to accept privacy recommendations that align with their learned profiles, leading to more privacy-protective settings.
  • The "Privacy Paradox" Persists: Even with tools, users may still face trade-offs where restrictive privacy settings break app functionality, highlighting the need for more granular control (e.g., purpose-based permissions).
  • Nudges Enhance Engagement but Not Necessarily Further Changes: While nudges increase awareness and engagement, they didn't lead to significant changes in settings already adopted based on PPA recommendations, suggesting the initial recommendations were quite stable.

Key Results

  • High Recommendation Adoption: 78.7% of the recommendations made by the PPA were accepted by users in the treatment group.
  • Stable Settings: Only 5.1% of the accepted recommendations were later re-allowed, indicating high precision and user satisfaction with the initial choices.
  • Reduced Configuration Effort: Participants in the treatment group made significantly more permission changes on the day they received recommendations (Day 3) compared to the control group, suggesting the PPA reduced the effort needed to configure settings.
  • User Comfort: Participants reported high comfort levels with the provided recommendations and their resulting privacy settings.
  • Profile Stability: 71.43% of participants in the treatment group remained in their initially assigned privacy profile throughout the study, indicating the profiles were a good fit.
  • Downsampling Validation: Learning privacy profiles from as few as 80-100 users yielded stable and reasonably accurate results (F-1 score > 0.70), suggesting the approach is robust even with limited initial data.

Limitations and Future Work

  • Rooted Device Dependency: The study required rooted Android devices, limiting the generalizability of the participant pool to tech-savvy users. Future work could explore methods for non-rooted devices.
  • Study Duration: The study was relatively short. Longer-term studies are needed to understand how privacy preferences evolve and how users adapt to personalized assistants over time.
  • Binary Permissions: The current permission model forces users into all-or-nothing choices. The authors suggest future work should explore purpose-based permission controls to allow for more nuanced decisions.
  • App Crashes: Some users re-allowed permissions because denying them caused app crashes, highlighting the need for better handling of such scenarios.
  • Improving Recommendations: The classifier could potentially be tuned for more aggressive recommendations, and access to larger datasets could further improve profile predictive power.
  • Beyond Mobile: The principles of personalized privacy assistants could be extended to other domains like web browsing and the Internet of Things (IoT).

Summary updated

BibTeX
@inproceedings{liu2016follow,
  author = {Bin Liu and Mads Schaarup Andersen and Florian Schaub and Hazim Almuhimedi and Shikun Zhang and Norman M. Sadeh and Yuvraj Agarwal and Alessandro Acquisti},
  title = {{Follow My Recommendations: A Personalized Privacy Assistant for Mobile App Permissions}},
  booktitle = {SOUPS},
  year = {2016},
  url = {https://openreview.net/forum?id=GZorIohz3J},
}
Figure