Research on Federated Sharing Methods for Massive Data in Blockchain

Data storage with the help of blockchain can ensure the transparency, non-tampering and autonomy of data information holders. However, the on-chain storage of massive data will seriously affect the performance of blockchain, and some private sensitive data and so on are not suitable for public storage in blockchain. To address the above problems, a trusted federated learning method based on local differential privacy mechanism, Loosely Coupled Local Differential Privacy Blockchain Federated Learning (LL-BCFL) is proposed for blockchain that realizes secure and efficient processing of massive user data. Firstly, a client selection mechanism is proposed and designed with the help of blockchain, which mainly includes two operations, namely, verification update and reputation calculation, to ensure the correctness and effectiveness of global model aggregation as well as the honesty and motivation of clients participating in training. Secondly, federated learning is used to realize the joint training of massive data distributed stored in each terminal device, so as to alleviate the phenomenon of “data silos” caused by privacy and security issues. In addition, a local differential privacy mechanism is designed in this method to solve the inference attack problem in the training process of federated learning. Finally, experiments are conducted on the MNIST dataset for both balanced and unbalanced datasets to verify the effectiveness of the proposed method LL-BCFL.