TL;DR
This paper introduces πBox, a novel application platform that shifts privacy protection from users and apps to the platform itself. By employing a device-and-cloud sandbox, specialized communication channels, and differential privacy, πBox enables a wide range of useful applications while significantly enhancing user privacy.
What Problem Does This Paper Address?
In today's digital landscape, mobile apps and web services often require access to sensitive user data (contacts, location, speech, etc.) to provide their core functionality. While users can grant or deny permissions, this system is flawed. Permissions are often coarse-grained, apps may request more than they need, and users are frequently overwhelmed, leading them to grant broad access without fully understanding the implications. This creates a significant privacy risk, as apps can misuse or leak user data. Existing solutions often force a stark trade-off between functionality and privacy, or rely on users to make complex privacy decisions.
Methods
πBox tackles this problem with a multi-pronged approach:
- Device-and-Cloud Sandbox: πBox extends the traditional app sandbox to span both the user's device and a cloud backend. This creates a unified, isolated environment for each user's app instance, ensuring that computations and data storage are confined.
- Specialized Communication Channels: Instead of allowing unrestricted network access, πBox provides five carefully designed channels for apps to interact with the outside world. These include:
- Private Vault: Per-user, per-app storage for sensitive data.
- Content Storage: Per-publisher storage for shared content (e.g., maps, articles).
- Aggregate Channel: For releasing privacy-preserving statistics to app publishers.
- Inbox: For receiving messages from publishers or other users.
- Sharing Channel: For explicit, user-authorized content sharing.
- Differential Privacy under Continual Observation: To enable publishers to collect useful statistics (like ad impressions or article views) without compromising individual user privacy, πBox adapts differential privacy algorithms. This is crucial for dynamic environments where data changes frequently. Specifically, they use "delayed-output counters" and techniques for "differentially private ranked lists" to improve accuracy compared to traditional methods.
Datasets
The paper evaluates πBox using:
- Micro-benchmarks: To measure the performance overhead of various πBox mechanisms.
- Web server trace: A 60-day trace from the 1998 World Cup website was used to evaluate the accuracy of the differential privacy mechanisms for reporting top-K accessed documents.
- University of Saskatchewan Web server trace: Used to compare the accuracy of delayed-output counters against basic counters for document access statistics.
- Google Play App Store Survey: A survey of top paid and free apps to assess the coverage and support of πBox for existing app functionalities.
No new benchmarks were introduced; rather, existing traces and real-world app categories were used for evaluation.
Key Insights and Ideas
- Shifting Responsibility: The core idea is to move privacy protection from the user and the app to the platform provider. This is a more robust approach, as users are often not equipped to make fine-grained privacy decisions, and apps can be untrusted.
- "Green," "Yellow," and "Red" Privacy Ratings: πBox categorizes apps based on the channels they use, providing a clear visual indicator of their privacy implications. "Green" apps are the most private, never exporting data. "Yellow" apps use aggregate channels with differential privacy guarantees. "Red" apps involve user-authorized sharing, carrying a higher risk.
- Balancing Act: πBox demonstrates that it's possible to achieve a useful balance between app functionality and user privacy, rather than forcing a strict trade-off.
- Practicality of Differential Privacy: The paper shows that differential privacy, particularly under continual observation, can be practically implemented in a working system to provide meaningful statistics with strong privacy guarantees.
Key Results
- Performance Overhead: For a simple static workload, πBox incurs a throughput reduction of about 50% and an overhead of 0.17 ms per operation. For more intensive workloads like SHA-256 hashing, the overhead is largely hidden. User isolation overhead was found to be insignificant across various workloads.
- Differential Privacy Accuracy: For reporting top-K accessed documents, the accuracy of the noisy lists improved significantly with higher privacy parameters (€). For example, the 8th-ranked item appeared in the noisy top-10 list 75% of the time with €=0.05 and 95% with €=0.2. Delayed-output counters showed a relative error of 12.9% for frequently accessed documents and 15.6% for infrequently accessed ones, compared to 19.6% and 83.1% for basic counters, respectively.
- App Coverage: A survey of top Google Play apps indicated that a significant portion (74% of paid apps and 67% of free apps, considering core functionality) could be supported by πBox, with most falling into the "green" or "yellow" categories.
Limitations and Future Work
- Limitations:
- Covert Channels: While πBox aims to prevent explicit leaks, the paper acknowledges that covert channels (e.g., timing) could still potentially be exploited, though this is an implementation-specific concern.
- Steganography: The sharing channel, while user-authorized, doesn't prevent apps from hiding private information within shared content using steganography.
- Absolute Privacy: The paper notes that absolute privacy is not achievable; some information leakage is inherent when data is released for utility.
- Trusted Platform: πBox relies on the platform provider being trusted, which is a fundamental assumption.
- Future Work:
- The authors suggest that πBox is well-suited for enterprise environments looking to enhance their app platforms or offer privacy-enhanced apps.
- They mention that conventional click-through ads could be supported with modifications, though this might involve privacy trade-offs.
- Further exploration of privacy-preserving ad auctions is also mentioned.
Summary updated