A data trading scheme based on blockchain and game theory in federated learning

Abstract

This paper proposes a federated learning scheme based on blockchain and game theory to address the main challenges in traditional federated learning models, including the risk of malicious user intrusion, imperfect data aggregation algorithms, and the problem of untraceable data. To solve the problems of malicious user intrusion and imperfect data aggregation algorithms, we introduce a verification game model and a weighted federated aggregation algorithm. By applying game theory principles to analyze users’ data verification behaviors, we generate a credit function and then integrate it into the weighted federated aggregation algorithm. This method significantly improves the data aggregation quality. In addition, our scheme takes advantage of the decentralized, transparent, and tamper-proof characteristics of blockchain to construct a verification body, realizing the functions of data flow recording and tracking. Experimental results show that when the participating nodes are only 75% and the training epoch is 7, the model accuracy reaches 98%, which is comparable to that of the full data model. In terms of data traceability, all data can be effectively tracked.