Towards heterogeneity-aware federated self-supervised learning via knowledge anchoring

Federated Self-Supervised Learning (FSSL) is a promising paradigm for extracting robust representations from decentralized unlabeled data. However, its effectiveness is often hindered by non-IID data distributions and label scarcity, which cause model divergence and limit generalization. In this paper, we propose Federated Self-Supervised and Global-Personalized Collaborative Learning (FedGP), a novel framework designed to bridge the gap between global knowledge integration and local client adaptation. The core of FedGP is the Collaborative Knowledge Anchoring (CKA) mechanism, which utilizes adaptive regularization to anchor shared global knowledge while enabling personalized refinement on local data. By dynamically balancing collaborative risks and local empirical losses via learnable coefficients, FedGP ensures stable convergence in heterogeneous environments. Extensive evaluations on multiple benchmarks, including a real-world private Flora dataset, demonstrate that FedGP consistently outperforms state-of-the-art FSSL methods. Our results confirm that FedGP achieves high-quality representation learning with significantly reduced communication overhead and annotation dependency, providing a scalable solution for privacy-preserving decentralized systems.