Efficient privacy-preserving federated logistic regression with poor-quality users

Federated learning has garnered substantial adoption across various domains due to its ability to address privacy protection and data silo challenges to a certain degree. The importance of logistic regression, as a fundamental machine learning algorithm, naturally extends to federated logistic regression. However, privacy inference attacks pose a significant threat, enabling servers to infer sensitive information from the updates uploaded by clients. Moreover, unavoidable errors in data collection and storage may result in low-quality updates, which in turn, reduce the overall effectiveness of the resulting model. Although several privacy-preserving federated logistic regression schemes have been proposed to safeguard privacy, few offer effective solutions to mitigate the adverse effects of low-quality data within a unified framework. To address these shortcomings, we propose a novel privacy-preserving federated logistic regression approach designed to mitigate the impact of low-quality users. Our method employs truth discovery techniques to reduce the influence of poor-quality updates on the global model, while utilizing a double-masking technique to ensure privacy. Theoretical analysis and experimental results validate that our approach successfully achieves both privacy preservation and robustness against the detrimental effects of low-quality users.