Cross-silo Federated Learning with Record-level Personalized Differential Privacy

Area of Science:

• Computer Science
• Machine Learning
• Cybersecurity

Background:

• Federated learning (FL) uses differential privacy to protect client data.
• Current FL privacy methods offer uniform protection, which may not suit individual data record needs.
• Personalized differential privacy in cross-silo FL remains an underexplored area.

Purpose of the Study:

• To introduce a novel framework for record-level personalized differential privacy in cross-silo federated learning.
• To address the challenge of determining optimal per-record sampling probabilities for personalized privacy budgets.
• To improve privacy preservation and performance in FL systems.

Main Methods:

• Developed the rPDP-FL framework with a two-stage hybrid sampling scheme (client-level and record-level).
• Introduced the Simulation-CurveFitting method to model the nonlinear relationship between sampling probability and privacy budget.
• Derived a mathematical model for per-record sampling probability (q) based on personalized privacy budget (ε).

Main Results:

• The proposed rPDP-FL framework effectively accommodates varying privacy requirements at the record level.
• Simulation-CurveFitting successfully identified the correlation between sampling probability and privacy budget.
• The derived mathematical model enables precise control over personalized privacy.
• Evaluations show significant performance gains compared to non-personalized privacy baselines.

Conclusions:

• Record-level personalized differential privacy is crucial for advanced federated learning applications.
• The rPDP-FL framework and Simulation-CurveFitting method offer a robust solution for personalized privacy in FL.
• This approach enhances data security and model performance by respecting individual data privacy needs.