"When and Where Do You Want to Hide?" - Recommendation of Location Privacy Preferences with Local Differential Privacy

Maho Asada, Masatoshi Yoshikawa, Yang Cao

Managing location privacy on modern smartphones has become a significant user experience challenge. As we move through the world, we are constantly prompted to share our location. While this data enables valuable services, it carries high risks of re-identification. Most users find it too tedious to manually set "hide" or "share" preferences for every hour of the day and every place they visit, leading to a gap between privacy awareness and actual practice.

Researchers at Kyoto University have proposed a system that recommends these privacy settings to users. By predicting when and where a user might want to stay "off the grid," the system reduces cognitive load while using Local Differential Privacy (LDP) to ensure the recommendation server never sees the user's true habits.

The Privacy-Utility Gap

Existing solutions for managing location privacy often rely on simple collaborative filtering or require manual input for every instance. These methods struggle to scale as the number of users and locations grows and often depend on a "trusted server" to handle sensitive data. In reality, users may not trust the service providers themselves.

Comparison of the proposed method with existing privacy-preserving recommendation techniques.

As shown in the comparison above, the authors move beyond traditional collaborative filtering by adopting Matrix Factorization (MF). This approach is more accurate for large-scale datasets and allows for the integration of mathematically rigorous privacy through LDP.

The Proposed Framework

The system combines two core techniques to balance utility and privacy:

  1. Matrix Factorization (MF): Similar to how streaming services predict movie preferences, MF predicts a user's privacy preference for a specific (time, location) pair by identifying patterns across the broader user base.
  2. Local Differential Privacy (LDP): This is the "privacy-by-design" component. Instead of sending raw data, the user’s device adds mathematical "noise" to their preferences and location visits before transmission. This ensures that even if the server is compromised, an attacker cannot determine a user's true location history.
System architecture of the LDP-based recommendation framework.

The architecture ensures that the "Client side" handles the sensitive user feature matrix computation, only sending perturbed gradients to the recommender system. The server then sends back an updated global model (V) which the client uses to generate local recommendations.

Key Insights and Results

To test the system, the researchers created a hybrid dataset combining LocShare (real-world privacy preferences) and Gowalla (massive trajectory data). Their analysis yielded several critical findings:

  • The 8-Hour Sweet Spot: Dividing the day into 8-hour blocks provided the best balance. It kept data dense enough for the algorithm to function while remaining fine-grained enough to capture distinct morning, afternoon, and night privacy needs.
  • High Accuracy with Noise: The system achieved a 90% probability of predicting accurate privacy values. Remarkably, the LDP-enabled model performed nearly as well as models using raw, unprotected data, even with strict privacy budgets (e.g., $\epsilon = 0.001$).
  • The 70% Threshold: The model reaches peak effectiveness when it knows a user's preference for approximately 70% of visited locations. Beyond this, the model risks "overlearning," which can decrease recommendation quality for new scenarios.

Limitations and Future Work

While the results are promising, the authors identify "temporal privacy leakage" as a remaining challenge, as location data is highly correlated over time. Future iterations aim to incorporate more complex spatial features—such as the influence of neighboring locations—and integrate the algorithm into live smartphone operating systems to study real-world user interaction.

Summary updated

Abstract

In recent years, it has become easy to obtain location information quite precisely. However, the acquisition of such information has risks such as individual identification and leakage of sensitive information, so it is necessary to protect the privacy of location information. For this purpose, people should know their location privacy preferences, that is, whether or not he/she can release location information at each place and time. However, it is not easy for each user to make such decisions and it is troublesome to set the privacy preference at each time. Therefore, we propose a method to recommend location privacy preferences for decision making. Comparing to existing method, our method can improve the accuracy of recommendation by using matrix factorization and preserve privacy strictly by local differential privacy, whereas the existing method does not achieve formal privacy guarantee. In addition, we found the best granularity of a location privacy preference, that is, how to express the information in location privacy protection. To evaluate and verify the utility of our method, we have integrated two existing datasets to create a rich information in term of user number. From the results of the evaluation using this dataset, we confirmed that our method can predict location privacy preferences accurately and that it provides a suitable method to define the location privacy preference.
BibTeX
@inproceedings{asada2019when,
  author = {Maho Asada and Masatoshi Yoshikawa and Yang Cao},
  title = {{"When and Where Do You Want to Hide?" - Recommendation of Location Privacy Preferences with Local Differential Privacy}},
  booktitle = {DBSec},
  year = {2019},
  url = {https://openreview.net/forum?id=KhLKdV2zNFjt},
}
Figure