The process of exchanging healthcare data introduces stringent requirements regarding users’ privacy. Federated learning (FL) is a novel model-sharing technique that aims to give additional privacy guarantees during machine learning process. Blockchain, as a form of distributed ledger technology, possesses the characteristic of trustworthiness; however, it is deficient in terms of computational capacity with a high-latency network due to its laborious consensus protocols. In this paper we present a distributed healthcare FL-based secure model sharing architecture to ensure healthcare data privacy and scalability. The solution relies on state channels technique to reduce on-chain transactions, contrast architecture latency, and reduce bandwidth consumption, alleviating the burden on the blockchain. State channels can be utilized to efficiently execute the tasks of federated learning models sharing and to solve the scalability problem.

The application of machine learning (ML) techniques to electronic health records (EHR) is gaining more and more attention as a method to extract valuable information that has the potential to enhance the decision-making process within the healthcare domain. A useful approach comes from the fed-erated learning (FL) scenario, which facilitates the decentralised training of machine learning models using datasets that are stored locally, hence eliminating the necessity of data aggregation on a central server. Federated learning also ensures data privacy because the federated devices do not share the actual data and store it locally. It becomes a useful tool when integrated with blockchain technology, which provides some properties such as immutability and traceability that are useful to enhance the security of such applications. With the growing use of IoT health care (loHT) devices, it is becoming challenging to manage them centrally and ensuring the health care data privacy. In this work, we propose an architecture to address the scalability issue related to the healthcare data management for federated learning networks with a sharding-based blockchain technique. We discuss some basic properties and report some results also coming from the implementation in Hyperledger Fabric.