Differentially private multivariate time series forecasting of aggregated human mobility with deep learning: Input or...

Area of Science:

• Computer Science
• Urban Planning
• Data Science

Background:

• Forecasting aggregated human mobility is crucial for urban planning.
• Ensuring individual privacy during data analysis is a significant challenge.
• Deep learning models offer powerful forecasting capabilities but raise privacy concerns.

Purpose of the Study:

• To investigate privacy-preserving methods for multivariate human mobility forecasting.
• To evaluate the effectiveness of differential privacy techniques in deep learning models for mobility data.
• To compare different recurrent neural network architectures under privacy constraints.

Main Methods:

• Implementation of differential privacy through gradient perturbation (differentially private stochastic gradient descent) and input perturbation.
• Comparison of four recurrent neural network models: Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Bidirectional-LSTM, and Bidirectional-GRU.
• Training and evaluation on a real-world multivariate mobility dataset.

Main Results:

• Differentially private deep learning models achieved performance comparable to non-private models.
• The performance loss was minimal, ranging between specific values (e.g., $$ and $$).
• Both gradient and input perturbation methods proved effective in preserving privacy with negligible accuracy impact.

Conclusions:

• Deep learning models trained with differential privacy offer a viable solution for accurate and private human mobility forecasting.
• The findings support the use of these methods in urban planning and decision-making processes.
• The study provides a publicly available dataset for further research in this domain.