A federated learning differential privacy algorithm for non-Gaussian heterogeneous data.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Federated learning faces challenges with multi-center heterogeneous data that often deviates from normal distributions.
• Non-normal data distributions in federated learning hinder model accuracy and robustness.

Purpose of the Study:

• To improve the DP-Fed-mv-PPCA model for federated learning with multivariate skewed normal client data.
• To develop a robust and privacy-preserving federated learning approach for heterogeneous data.

Main Methods:

• Utilized a Bayesian framework to define prior distributions for local parameters.
• Employed expectation-maximization and pseudo-Newton algorithms for robust parameter estimation.
• Integrated clipping and differential privacy algorithms to ensure model parameter solutions and privacy.

Main Results:

• The improved DP-Fed-mv-PPCA model demonstrates effectiveness in handling non-normally distributed data.
• Achieved robust parameter estimation and differential privacy guarantees.
• Validated the model's performance on both synthetic and real-world Internet of Vehicles data.

Conclusions:

• The proposed federated learning model effectively addresses challenges posed by multi-center heterogeneous data with skewed distributions.
• The method provides a robust and privacy-preserving solution for federated learning applications.
• The model shows promise for real-world applications, particularly in domains like the Internet of Vehicles.