Local Model Quality Control Method Based on Credit Mortgage for Enterprise Credit Evaluation

Abstract

Establishing enterprise credit evaluation models requires high-quality data from enterprises. Considering privacy concerns of enterprises, blockchain and federated learning architecture become an effective solution. However, this solution cannot avoid malicious training behaviors of enterprises in order to obtain high profits, resulting in the degradation of global model performance. To cope with the problems above, this paper proposes a local model quality control method based on credit mortgage value. Specifically, each enterprise submits a portion of its credit value as a credit mortgage to participate in federated learning, then each enterprise obtains rewards or penalties according to the quality of the enterprises local model combined with the credit mortgage value. Simulation results show that our method can improve the performance of the global model with malicious behavior on the clients. For example, when the malicious training probability of the client increases to 30%, the credit mortgage value required by the client to participate in federated learning becomes 6.2 times the original value, and revenues turn to be negative. Severe negative revenues can form an effective control over the quality of the local model. So our method can realize effective control of local model Quality.