查看文章

Fairness across different demographic groups is an essential criterion for face-related tasks, Face Attribute Classification (FAC) being a prominent example. Simultaneously, federated Learning (FL) is gaining traction as a scalable paradigm for distributed training. In FL, client models trained on private datasets get aggregated by a central aggregator. Existing FL approaches require data homogeneity to ensure fairness. However, this assumption is restrictive in real-world settings. E.g., geographically distant or closely associated clients may have heterogeneous data. In this paper, we observe that existing techniques for ensuring fairness are not viable for FL with data heterogeneity. We introduce F3, an FL framework for fair FAC under data heterogeneity. We propose two methodologies in F3, (i) Heuristic-based and (ii) Gradient-based, to improve fairness across demographic groups without requiring data …