A Neural Approach to Spatio-Temporal Data Release with User-Level Differential Privacy.

Area of Science:

• Data privacy
• Machine Learning
• Spatio-temporal data analysis

Background:

• Publicly released aggregate location data from companies like Meta and Google support applications in transportation, public health, and urban planning.
• Differential privacy (DP) is the standard for protecting individual location data, but current methods reduce data utility under user-level privacy (multiple reports per individual).
• Existing "data-for-good" initiatives often use high privacy budgets (ε=10-100), compromising user privacy.

Purpose of the Study:

• To propose a novel approach for private and accurate release of spatio-temporal data, addressing the limitations of current DP methods for user-level privacy.
• To improve the utility of differentially private location data while maintaining robust privacy guarantees.

Main Methods:

• Utilizing variational auto-encoders (VAEs), a type of neural network, to leverage pattern recognition capabilities.
• Applying VAEs to reduce noise introduced by DP mechanisms, thereby enhancing data accuracy.
• Integrating DP with VAEs to satisfy privacy requirements while improving data utility.

Main Results:

• The proposed VAE-based approach significantly increases the accuracy of released spatio-temporal data compared to existing benchmarks.
• The method effectively reduces the noise inherent in DP mechanisms when handling multiple data points per user.
• Experimental evaluations on real-world datasets demonstrate the superiority of the VAE-enhanced DP approach.

Conclusions:

• The novel VAE-based method offers a superior solution for private spatio-temporal data release, balancing accuracy and privacy.
• This approach overcomes the utility-privacy trade-off limitations of traditional DP methods in user-level privacy scenarios.
• The findings suggest a promising direction for enhancing the practical application of "data-for-good" initiatives.