Secure and decentralized federated learning framework with non-IID data based on blockchain.

Area of Science:

• Computer Science
• Artificial Intelligence
• Blockchain Technology

Background:

• Federated learning facilitates collaborative model training without data sharing, but struggles with non-independent and identically distributed (non-IID) data.
• Non-IID data distributions across organizations significantly challenge traditional federated learning performance and model accuracy.

Purpose of the Study:

• To propose a hierarchical federated learning framework leveraging blockchain technology to enhance non-IID data training.
• To improve data privacy, security, and overall federated learning performance in decentralized environments.

Main Methods:

• Developed a blockchain system to create a global shared pool, reducing local data non-IID degree and enhancing model accuracy.
• Utilized smart contracts for decentralized model distribution, collection, and aggregation on a main blockchain.
• Trained Multi-Layer Perceptron (MLP) and Convolutional Neural Network (CNN) models on MNIST, Fashion-MNIST, and CIFAR-10 datasets.

Main Results:

• The proposed framework significantly improved the accuracy of decentralized federated learning models.
• Demonstrated effectiveness in handling non-IID data across multiple benchmark datasets.
• Validated the feasibility and performance enhancements of the blockchain-based hierarchical federated learning approach.

Conclusions:

• The hierarchical federated learning framework effectively addresses non-IID data challenges in decentralized settings.
• Blockchain integration enhances privacy, security, and performance in federated learning.
• The proposed method offers a robust solution for collaborative machine learning with heterogeneous data distributions.