Data Compliance Utilization Method Based on Adaptive Differential Privacy and Federated Learning.

Area of Science:

• Computer Science
• Cybersecurity
• Data Science

Background:

• Federated learning (FL) trains models collaboratively without sharing raw data, offering inherent privacy benefits.
• However, FL systems remain vulnerable to inference attacks that can expose sensitive user data through intermediate model parameters.
• Existing privacy-preserving methods may not adequately balance model accuracy with robust security against sophisticated attacks.

Purpose of the Study:

• To propose a novel Adaptive Differential Privacy Blockchain Federated Learning (ADP-BCFL) method.
• To enhance the security of federated learning against data inference attacks.
• To ensure compliant and secure utilization of distributed data while maintaining high model performance.

Main Methods:

• Implemented a blockchain framework for secure storage and querying of aggregated user data.
• Developed an adaptive differential privacy (DP) mechanism to dynamically adjust noise levels based on parameter characteristics.
• Integrated DP into the federated learning process to control information leakage and mitigate inference risks.

Main Results:

• The ADP-BCFL method demonstrated effective prevention of sensitive data inference attacks.
• Adaptive DP successfully balanced the introduction of noise to protect privacy without significantly degrading global model accuracy.
• Validation on MNIST, Fashion MNIST, and spatiotemporal datasets confirmed the method's efficacy.

Conclusions:

• The ADP-BCFL method provides a robust solution for secure and private federated learning.
• Blockchain integration ensures data integrity and secure access to aggregated information.
• The adaptive DP mechanism offers a flexible approach to privacy preservation in FL, crucial for sensitive distributed datasets.