A Federated Learning Scheme Based on Reputation Evaluation and Blockchain

Federated learning is an emerging distributed machine learning paradigm, where multiple data sources collaborate to train a model without sharing the original data, safeguarding user data privacy. Nevertheless, users, whether intentionally or unintentionally, may contribute low-quality data that can impact model quality or infer user information from the model parameters, potentially leading to privacy breaches. Selecting appropriate participants for training is still a challenge needing resolution. This paper introduces a federated learning scheme called FREB, which leverages a reputation evaluation mechanism and blockchain technology. The foundation of this scheme is blockchain, and it incorporates reputation as an indicator of participant reliability. Reputation is computed using a multi-weight subjective logic model. Before the training process begins, participants are chosen based on their reputation. Furthermore, the paper presents a novel model aggregation approach. By injecting noise into the uploaded model parameters, the scheme ensures that potential attackers cannot glean sensitive data information from the parameters. To validate the efficacy of FREB, the authors conducted simulation experiments using real datasets. The results of these experiments demonstrate that FREB is capable of thwarting malicious node attacks and promoting high-quality collaborative learning. It achieves this while preserving data privacy and maintaining model accuracy.