Hierarchical Aggregation for Numerical Data under Local Differential Privacy.

Area of Science:

• Computer Science
• Data Privacy
• Statistical Analysis

Background:

• Centralized differential privacy models require trusting data collectors.
• Local differential privacy (LDP) addresses this by enabling local data perturbation.
• Existing LDP methods often lack flexibility for data with varying sensitivity ranges.

Purpose of the Study:

• To propose a hierarchical aggregation framework for numerical data under LDP.
• To address the need for differentiated privacy protection based on data ranges.
• To improve the accuracy of statistical analysis in LDP settings.

Main Methods:

• A hierarchical aggregation framework is proposed where data is disturbed locally based on assigned privacy levels.
• An aggregator further perturbs data to convert lower-level privacy data to higher levels.
• The framework is integrated with mini-batch stochastic gradient descent for linear regression.

Main Results:

• Theoretical analysis confirms the framework adheres to LDP requirements and provides unbiased mean estimation.
• Experiments on synthetic and real datasets demonstrate superior accuracy compared to existing methods.
• Improved performance is observed in both mean estimation and linear regression tasks.

Conclusions:

• The proposed hierarchical framework effectively enhances privacy protection for numerical data with diverse ranges under LDP.
• This approach offers a more accurate and flexible solution for statistical analysis in LDP environments.
• The method shows significant potential for real-world applications requiring robust data privacy.