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

arXiv: https://arxiv.org/abs/2208.09744 · HTML: https://mlsanity.com/paper/2208.09744/

## Key facts

- Cited by 5 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 20 Aug 2022.

## Built on by (most cited first, 5 of 5)

- [Adaptive Hypergraph Network for Trust Prediction](https://mlsanity.com/paper/2402.05154/) (2024)
- [Large Language Model Powered Intelligent Urban Agents: Concepts, Capabilities, and Applications](https://mlsanity.com/paper/2507.00914/) (2025)
- [RLER-TTE: An Efficient and Effective Framework for En Route Travel Time Estimation with Reinforcement Learning](https://mlsanity.com/paper/2501.15493/) (2025)
- [BiasBuster: a Neural Approach for Accurate Estimation of Population Statistics using Biased Location Data](https://mlsanity.com/paper/2402.11318/) (2024)
- [Differentially private synthesis of Spatial Point Processes](https://mlsanity.com/paper/2502.18198/) (2025)

## Abstract (arXiv)

Several companies (e.g., Meta, Google) have initiated "data-for-good" projects where aggregate location data are first sanitized and released publicly, which is useful to many applications in transportation, public health (e.g., COVID-19 spread) and urban planning. Differential privacy (DP) is the protection model of choice to ensure the privacy of the individuals who generated the raw location data. However, current solutions fail to preserve data utility when each individual contributes multiple location reports (i.e., under user-level privacy). To offset this limitation, public releases by Meta and Google use high privacy budgets (e.g., ε=10-100), resulting in poor privacy. We propose a novel approach to release spatio-temporal data privately and accurately. We employ the pattern recognition power of neural networks, specifically variational auto-encoders (VAE), to reduce the noise introduced by DP mechanisms such that accuracy is increased, while the privacy requirement is still satisfied. Our extensive experimental evaluation on real datasets shows the clear superiority of our approach compared to benchmarks.