Toward Fair Federated Edge Learning Through Prototype-Guided Distributed Adversarial Networks

Abstract

Federated learning (FL) is a privacy-preserving machine learning (ML) paradigm that can learn models from distributed datasets owned by mobile terminals (MTs). However, ML models usually contain bias on some user groups/sensitive attributes (e.g., gender and race), which poses additional challenges on FL in real-world applications. Some techniques have been proposed to train fair models in FL, but they often require MTs to share sensitive information about private datasets. Moreover, existing methods neglect a crucial and practical scenario, i.e., sensitive attributes are non-independent and identically distributed across MTs, where some MTs can only observe partial sensitive attributes, thereby generating a biased local model. In this paper, we propose a fairness-aware FL framework via prototype-guided distributed adversarial networks, namely FAFL, to achieve target fairness requirements. Specifically, in FAFL, each MT is equipped with a discriminator to distinguish model bias based on latent representations from deployed models and is required to train a multi-objective model including minimizing the classification error and the model bias locally. Meanwhile, considering the heterogeneity of sensitive attributes, FAFL aggregates the uploaded attribute-specific prototypes of local discriminators into a global prototype that is utilized to regularize the local training. Experiments on three real-world datasets demonstrate that our method can effectively improve model fairness even when the distribution of sensitive attributes across MTs is extremely unbalanced.