Privacy-preserving federated feature selection with differential privacy

There is an urgent need to perform effective feature selection in distributed environments while preserving data privacy. In this paper, a new federated feature selection framework is developed to protect the privacy of input features held by multiple distributed clients, with applications in engineering systems where secure and efficient feature selection is critical in distributed environments. The proposed framework is based on federated learning and differential privacy techniques for distributed environments. The distributed clients send the noisy features’ values to the server preserving the privacy. The server then aggregates these noisy features’ values for further computations and feature selection. The performance of the proposed framework is evaluated against a centralized scenario where feature selection occurs centrally. Comparative analysis involves inputting the selected features into various machine learning models employing various evaluation metrics. The simulation results indicate comparable performance between the proposed federated approach and the centralized method. To further compare performance, a new method, ’Rank of Features’ is developed in this paper that evaluates similarity between features selected by the proposed framework and centralized method. The results of this analysis also demonstrate strong similarity between the two approaches. Further, privacy analyses are conducted in detail that include protection against reconstruction and membership inference attacks demonstrating robust preservation against data leakage and unauthorized inference of sensitive information.