A Blockchain‑Based Trusted Federated Learning Architecture: Overview, Progress, and Future Directions

This article introduces a blockchain‑based trusted federated learning architecture, beginning with an overview of privacy computing, its legal and regulatory background in China and abroad, and the market outlook for privacy‑preserving technologies.

It then discusses the integration of blockchain and privacy computing, explaining why blockchain alone lacks privacy guarantees and how privacy‑preserving computation (secure multi‑party computation, federated learning, and confidential computing) complements blockchain to enable secure data sharing.

Several industry cases are presented, including WeBank’s WeDPR‑PPC platform, AntChain’s FAIR network, QuarkChain’s financial data sharing platform, and government tax data platforms, illustrating practical deployments of blockchain‑plus‑privacy computing.

The paper reviews federated learning fundamentals, its workflow (global model distribution, local training, gradient aggregation), and notable applications such as Google Gboard, Apple Siri, Intel brain‑tumor detection, and Nvidia mammogram analysis.

Key challenges of federated learning—statistical heterogeneity, system heterogeneity, communication cost, and security/robustness—are identified, motivating a decentralized federated architecture.

A novel framework called CMFL (Committee‑based Decentralized Federated Learning) is proposed. CMFL organizes nodes into training clients, committee clients, aggregation clients, and idle clients, and defines four functional modules: scoring, selection, election, and a committee consensus protocol (CCP) based on PBFT.

The technical workflow includes smart‑contract‑driven model initialization, random committee selection, local training, gradient submission to committee nodes, scoring, on‑chain aggregation, and automatic election of new committee members for each round.

Experimental evaluation on two datasets demonstrates CMFL’s robustness against Byzantine attacks, superior accuracy under both attack and non‑attack scenarios, and communication efficiency via an asynchronous ECMFL variant.

Finally, the article outlines future directions, emphasizing large‑scale data circulation, data assetization, and industry transformation in government, finance, and healthcare, while acknowledging remaining challenges in privacy‑blockchain integration.