FeDa4Fair: Client-Level Federated Datasets for Fairness Evaluation
Researchers have introduced FeDa4Fair, a novel benchmarking framework designed to address the "illusion of fairness" in Federated Learning (FL). While FL enables collaborative model training with privacy preservation, global models often appear fair on average while masking persistent discrimination at the client level. Existing solutions typically fail to handle complex, heterogeneous bias scenarios, such as attribute-bias and value-bias across different clients. To resolve this, the team developed a library for creating datasets tailored to evaluate fair FL methods under these challenging conditions. The project includes three main contributions: the FeDa4Fair library itself, a standardized benchmark suite generated by the library, and ready-to-use functions for evaluating fairness outcomes. This initiative aims to support more robust and reproducible research in algorithmic fairness within distributed learning environments. By stress-testing fairness methods against realistic conflicting biases, FeDa4Fair provides essential tools for developers and researchers to ensure equitable AI performance across diverse user groups, moving beyond simplistic binary attribute mitigation strategies currently prevalent in the field.
Wire timeline
FeDa4Fair: Client-Level Federated Datasets for Fairness Evaluation
Researchers have introduced FeDa4Fair, a novel benchmarking framework designed to address the "illusion of fairness" in Federated Learning (FL). While FL enables collaborative model training with privacy preservation, global models often appear fair on average while masking persistent discrimination at the client level. Existing solutions typically fail to handle complex, heterogeneous bias scenarios, such as attribute-bias and value-bias across different clients. To resolve this, the team developed a library for creating datasets tailored to evaluate fair FL methods under these challenging conditions. The project includes three main contributions: the FeDa4Fair library itself, a standardized benchmark suite generated by the library, and ready-to-use functions for evaluating fairness outcomes. This initiative aims to support more robust and reproducible research in algorithmic fairness within distributed learning environments. By stress-testing fairness methods against realistic conflicting biases, FeDa4Fair provides essential tools for developers and researchers to ensure equitable AI performance across diverse user groups, moving beyond simplistic binary attribute mitigation strategies currently prevalent in the field.
cs.AI updates on arXiv.org