Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Feature selection in the Federated Machine Learning setting

Christofilogiannis Ioannis

Full record

---

Summary

This thesis presents two main contributions to advance research in Federated Learning (FL): Feature Election and the FLEx framework. Feature Election is a novel federated Feature Selection algorithm that enables conventional Feature Selection (FS) methods to operate in horizontal federated settings without altering their core logic. The algorithm leverages client-generated vote vectors with preference scores while preserving data privacy, using a freedom degree parameter to control selection granularity. The second contribution, FLEx (Federated Learning Exchange), is a comprehensive framework that combines C++’s network performance with Python’s machine learning capabilities through Cython integration, secured by both symmetric and asymmetric encryption. This framework compares favorably with competing solutions based on evaluation metrics from a recent survey. Experimental validation across five datasets using three Machine Learning (ML)model types demonstrates that these contributions significantly reduce communication overhead with model parameter size reductions across all experiments (up to 93.4%), while maintaining or improving model performance and reducing noise, overfitting and computational cost. Integration of the Feature Election algorithm with the Flower framework achieved model size reductions up to 67.7%, while feature augmentation experiments confirmed robustness in high-dimensional spaces. Feature Election in FLEx establishes a new paradigm for network-efficient FL in bandwidth-constrained scenarios where data privacy is paramount.

Available Files

Services