BiFCL: A Bi‐Level Federated Contrastive Learning Framework With Auxiliary Model‐Guided Alignment

Abstract

Federated learning (FL) enables collaborative model training over distributed clients without sharing raw data, but its performance can degrade severely under non‐independent and non ‐ identically distributed (non‐IID) client data. To address this challenge, we propose BiFCL, a Bi‐Level Federated Contrastive Learning framework that improves global representation alignment through a lightweight server‐side auxiliary model trained on a small public dataset. BiFCL enforces bidirectional semantic consistency: on the client side, each local model is regularized by contrastively aligning its representations to the auxiliary model; on the server side, the auxiliary model is further updated via contrastive alignment with the evolving global model, preventing stale guidance and forming a closed‐loop regularization mechanism. We provide convergence guarantees for BiFCL under standard smooth non‐convex assumptions. Extensive experiments on five benchmarks (MNIST, EMNIST, SVHN, CIFAR‐10 and CIFAR‐100) demonstrate that BiFCL consistently outperforms strong FL baselines in both final accuracy and convergence speed. Overall, BiFCL provides an effective representation‐alignment strategy for heterogeneous federated learning when a small, diverse, and semantically relevant public, proxy or privacy‐preserving synthetic auxiliary source is available. We also discuss practical auxiliary‐data construction strategies, normalized communication cost and limitations under severe auxiliary‐domain mismatch and larger‐client deployments.