A Secure and Private Distributed Machine Learning Training Method for Smart Home Healthcare Systems Based on Block Chains and Decentralized Connectivity Learning

Abstract

The proposed solution is the PS-DFL (Private and Secure Decentralized Federated Learning) method, a secure and private decentralized federated learning (DFL) method based on blockchain technology, applied to smart home healthcare systems. The method utilizes a clustering-verification mechanism, where noisy updates are clustered using the K-means algorithm and verified using the Multi-Krum algorithm to efficiently handle heterogeneous data. The solution employs two layers of privacy protection, including: using pre-committed differential noise (DP noise) in the verification phase, and applying light differential privacy (light DP) in the synthesis phase. The aggregation process is performed securely through Verifiable Secret Sharing (VSS) based on the Shamir Secret Sharing scheme combined with Polynomial Commitment, allowing for the recovery of the aggregate update without revealing information about individual client devices. Furthermore, the system utilizes a dynamic staking update mechanism to select committees, including a Noising Committee, a Verification Committee, and an Aggregation Committee, to maintain decentralization and limit the concentration of stakes. This ensures privacy protection, enhances resilience against Byzantine attacks, and makes the method suitable for deployment on networked medical devices in a home environment.