An improved blockchain-based multi-region Federated Learning framework for crop disease diagnosis

Abstract

Plant Electronic Medical Records (PEMRs), containing crop disease and environmental data, provide a novel approach for disease diagnosis. However, training a federated learning (FL) model with PEMRs distributed across multiple devices poses challenges, such as opaque aggregation processes, risks of Byzantine faults, and high communication overhead. In this paper, we develop a blockchain-based multi-region federated learning (BMRFL) framework for crop disease diagnosis, incorporating the consortium blockchain technology to ensure that the process is both verifiable and resistant to attacks. We introduce the Musig2 Signature-based Practical Byzantine Fault Tolerance (M2SPBFT) protocol, which leverages the Musig2 algorithm to improve efficiency by reducing communication overhead and streamlining the verification process. Furthermore, we develop a aggregation strategy that boosts the global model's accuracy in diagnosing crop diseases. We constructed PEMR datasets with 23,702 samples from Beijing Plant Clinics to validate the BMRFL. Extensive experiments revealed that the BMRFL framework improved Byzantine fault resistance, lowered consensus communication overhead, and enhanced diagnostic accuracy across districts, achieving a 10.44 % accuracy increase in Haidian over previous methods. These results demonstrate the effectiveness and security of BMRFL in crop disease diagnosis, suggesting its potential for related diagnostic applications.