FedTail-DT: A Dual-Teacher Framework for Long-Tailed Heterogeneous FL with CLIP Prototypes and Adaptive Aggregation

Abstract

Federated learning enables decentralized model training while preserving data privacy, showing significant potential in sensitive fields such as healthcare and finance. However, real-world client data often exhibits both heterogeneity and long-tailed label distributions, leading to class bias and knowledge gaps that limit global model generalization. This paper proposes FedTail-DT, a novel approach for long-tailed heterogeneous federated learning, featuring three key innovations: (1) a dual-teacher–single-student architecture that integrates a pre-trained CLIP model as a semantic prior teacher and a globally aggregated model as an iterative optimization teacher to enhance tail-class perception; (2) CLIP text-prototype contrastive learning that uses text prototypes as semantic anchors to counteract feature bias caused by long-tailed distributions; (3) a multi-dimensional client scoring mechanism that dynamically calibrates aggregation weights based on data volume, category completeness, and update consistency. Experiments on federated benchmark datasets with various imbalances, for example, under extreme imbalance (factor 0.01), show that FedTail-DT outperforms existing SOTA methods by 4.21% on CIFAR-10-LT and 3.14% on CIFAR-100-LT, demonstrating its effectiveness in handling both data heterogeneity and long-tailed problems.