Dual-Model Active Federated Learning Based on Homomorphic Encryption

Abstract

With the rapid growth of data and the improvement of privacy protection requirements, federated learning and homomorphic encryption technology has become an important means to solve data cooperation and privacy protection. This paper proposes a dual-model active federated learning method based on homomorphic encryption, which aims to integrate uncertainty and diversity (representativeness) data selection and protect data privacy. This method utilizes two types of clients. One type of client uses uncertainty-based query strategies to select the data to be marked by evaluating the uncertainty of prediction results; the other is based on diversified query strategy, which selects the data to be marked by calculating the difference between the marked data and the marked data, and uses the data for local model training and parameter updating. Then, the model parameters of the two types of clients are encrypted using homomorphic encryption technology, and are safely uploaded to the server for model parameter aggregation. The server realizes comprehensive model update by aggregating the model parameters of the two types of models, and uses homomorphic encryption technology to protect data privacy. Through iterative training and model updating, we achieve an efficient active learning process that improves model performance while ensuring data privacy.