A Blockchain-Assisted Trusted Federated Learning for Smart Agriculture

Abstract Smart agriculture promises to alleviate the burden of climate risks on crop production by leveraging machine learning tasks. These tasks act as a decision support instrument for making well-informed choices by stakeholders in the agricultural value chain. Currently, predictive models in smart agriculture demand a centralized collection of diverse data, fragmented across multiple information systems leading to a single point of failure. The application of the Federated Learning (FL) technique restricts the movement of raw data and trains the model at the data source. However, the FL approach does not ensure trust factors like privacy, authentication, data provenance, transparency and traceability. To address this, a decentralized federated learning framework built on blockchain can be a potential solution. In this study, we introduce a blockchain-based framework called AgriFLChain for trusted federated learning in the context of smart agriculture. We focus on crop yield prediction as an illustrative use case, initially discussing centralized deep learning models (ResNet-16, ResNet-28, CNN-DNN, and CNN-LSTM). We then detail the authentication and data provenance mechanisms for federated learning participants, utilizing blockchain-based Decentralized Identifiers (DIDs) and Verifiable Credentials (VCs). We implement these models using vanilla federated learning and Differential Privacy (DP) federated learning approaches, achieving transparency and traceability through smart contracts by recording metadata of model updates into the blockchain. Finally, detailed evaluation demonstrates that AgriFLChain achieves comparable efficiency to centralized models while maintaining scalability in blockchain transactions for higher data volumes.