Research On Data Security Sharing Based On Blockchain And Federated Learning

With the advent of the big data era,data sharing has become a hot topic.Traditional centralized data sharing faces challenges in data security,data privacy,and processing efficiency.To address these issues,Google proposed a privacy-protected distributed machine learning framework called Federated Learning(FL)in 2016,which enables efficient and secure data sharing without accessing local data.However,FL-based data sharing still faces challenges regarding data heterogeneity and incentivize sharing.This paper aims to investigate two challenges and provide solutions.To address the challenge of data heterogeneity in FL-based data sharing,this paper proposes a novel algorithm called Clustered Federated Learning with Scale Constraint(CFLSC).CFLSC utilizes the similarity in data distributions across devices to form training clusters,reducing the impact of data heterogeneity.Compared with the previous CFL algorithm,CFLSC incorporates pre-training and a scale-constrained clustering algorithm to limit the size of device clusters,preventing device accumulation and overfitting,thus enhancing the efficiency and stability of device cluster initialization.Furthermore,CFLSC introduces an adaptive cluster guidance strategy to dynamically capture changes in data distribution among device clusters,reducing misjudgments of outdated cluster guidance gradients for new device clusters.Additionally,by combining momentum optimization,the potential of cluster models is fully unleashed.Finally,CFLSC is evaluated on three federated heterogeneous datasets.Experimental results demonstrate accuracy improvements of over 15%compared to traditional FL algorithms.When compared to the baseline CFL algorithm,CFLSC achieves an average accuracy improvement of nearly 4%.To address the challenge of incentive sharing in FL-based data sharing,this paper proposes an innovative incentive sharing scheme called Smart Contract and Model Value Transfer Incentive(SC-MVTI).it explicitly defines the bonus,which originates from the reward of the data requester,and the model bail from the low-quality data provider.It establishes a training pattern led by the data requester,and resolves the dilemma of high contribution but low return by employing a prepayment model bail,thus achieving secondary distribution of model benefits to the data providers.Additionally,SC-MVTI utilizes smart contracts and Top-k sparsification techniques to ensure the safe operation of the incentive mechanism and maintain the legitimate rights of both data providers.Finally,MVTI is validated on four Mnist partitioning scenarios.Experimental results demonstrate that MVTI outperforms other benchmark algorithms in terms of model accuracy,correctness of contribution evaluation,and fairness of distribution.Finally,based on the two aforementioned works,a data sharing system based on blockchain and federated learning has been developed.it aims to facilitate secure data sharing among organizations and institutions,promoting the establishment of broader data collaborations.