Enforcing group fairness in privacy-preserving Federated Learning

Abstract

With the expansive application of Federated Learning (FL) in critical domains such as healthcare and financial services, ensuring group fairness in FL has become imperative. The intrinsic distributed architecture of FL, while enhancing training efficiency, introduces significant challenges in maintaining group fairness due to client data privacy concerns and heterogeneous data distribution. For this reason, we present a suite of innovative strategies to navigate these challenges. First, we construct an approximate global dataset to achieve global group fairness. It was accomplished by collecting statistical information from clients and generating an approximate global dataset that circumvents the direct sharing of raw data. Moreover, we present the Local Sensitive Hashing-Categorical Attribute Data Completion (LSH-CADC) method, which improves the precision and efficiency of the generation of the approximate global dataset. In addition, an adaptive fairness-weighted aggregation (AFA) algorithm was developed to enhance the model’s group fairness. By integrating these methodologies within the classical privacy-preserving FL method HybridAlpha, we get a new framework, enforcing group fairness in privacy-preserving Federated Learning (GFL). GFL not only ensures privacy but also solidifies group fairness. Through extensive experimentation in various cross-dataset/domain non-IID scenarios, GFL demonstrated exceptional performance and group fairness metrics, surpassing existing benchmarks and highlighting its significant potential to promote group fairness.