Towards Group Fairness via Semi-Centralized Adversarial Training in Federated Learning

As federated learning increasingly performs better on tasks of decision-making scenarios such as medical care or commercial area, there have been concerns about discrimination against certain populations with sensitive attributes (e.g., race, gender). In this work, we propose to improve group fairness with semi-centralized adversarial training. And we adopt Variational AutoEncoder (VAE) for federated learning scenarios to generate adversarial samples. We keep VAE decoder at server side and leave encoder at client side to encode local samples into feature dimensions for transmitting, which ensures the privacy of user data. Our proposal further performs sensitive attribute alignment to improve group fairness. Our experimental evaluation shows that our approach outperforms the state-of-the-art federated learning frameworks in terms of group fairness and communication resource consumption.