A Two-Stage Differential Privacy Scheme for Federated Learning Based on Edge Intelligence.

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Security

Background:

• Federated learning (FL) enables collaborative model training without sharing raw data, but protecting sensitive information remains a challenge.
• Edge intelligence offers distributed processing capabilities, crucial for privacy-preserving FL frameworks.
• Existing FL methods may not sufficiently address data privacy concerns, necessitating advanced protection mechanisms.

Purpose of the Study:

• To propose a novel two-stage differential privacy (DP) framework for federated learning (FL) integrated with edge intelligence.
• To enable adjustable privacy preservation levels based on data sensitivity.
• To ensure data privacy and model security in distributed learning environments.

Main Methods:

• Implemented a two-stage DP framework: user-terminal feature perturbation using randomized response and edge-server local model noise addition.
• Utilized an end-edge-cloud architecture for federated learning.
• Employed a Bi-directional Long-Short-Term Memory (BiLSTM) neural network for classification on an electrocardiogram (ECG) dataset.

Main Results:

• The proposed framework demonstrated effective privacy preservation through adjustable perturbation and noise addition.
• Experimental results showed that the framework achieves good training accuracy and convergence on an ECG dataset.
• Analysis confirmed the impact of varying privacy budgets and parameters on model performance.

Conclusions:

• The developed two-stage DP framework successfully balances privacy protection and model performance in federated learning.
• The framework offers a flexible approach to privacy preservation, adaptable to different data sensitivities.
• This research contributes a robust solution for secure and efficient distributed machine learning.