Blockchain-Assisted Semi-Centralized Federated Learning Framework

With the development of network technology, building a trusted new-generation information management system is necessary. Blockchain technology provides a decentralized, transparent, and tamper-proof distributed base. On the other hand, with the development of artificial intelligence technology, data islands have been a common issue in the field of network data computing. The distrust among developers has made it difficult to jointly utilize all parties’ data for collaborative training. Although federated learning provides data privacy protection, there are still hidden dangers in server-side security. The traditional methods replace the server in the federated learning framework with a blockchain system to provide a tamperproof global model database. However, this approach does not utilize all available network connections in the Internet of things scenario and lacks a block structure design for federated learning tasks. We propose a blockchain-assisted semi-centralized federated learning framework. Starting from the requirements of the Internet of things scenario, our approach constructs a semi-centralized Internet of things structure and utilizes all trusted network connections to support federated learning tasks. At the same time, our approach constructs a tamper-proof model database for untrusted and remote clients through blockchain technology. Compared with traditional blockchain federated learning frameworks, our approach has a smaller communication overhead and better universality. The framework includes two major designs. The semi-centralized federated learning framework reduces the communication overhead brought by aggregation through trusted connections between clients, and stores client models through blockchain for aggregation on remote or untrusted clients to improve the universality and performance of local models. The design of blockchain blocks for federated learning tasks can support the needs of underlying federated learning training. Experiments have shown that this framework has an accuracy improvement at least 8% compared with traditional federated learning algorithms on multiple datasets, and significantly reduces the communication overhead caused by the waiting aggregation process between clients, providing guidance for the deployment of blockchain federated learning systems in practical scenarios.