ScholarGate
Get started
Machine learningMachine learningMachine learningAlgorithm

Semi-supervised Federated Learning

Also known as: SSL-FL, federated semi-supervised learning, FSSL, semi-supervised distributed learning

OriginatorJeong, W. et al. / multiple independent groupsYear2020Sources2Related methods8

Semi-supervised federated learning (SSFL) trains a shared model across many decentralized clients — each holding private data — when only a subset of clients or a subset of local samples carry labels. It combines the privacy-preserving coordination of federated learning with the label-efficiency of semi-supervised techniques such as pseudo-labeling and consistency regularization, enabling strong model quality without centralizing sensitive data.

Key highlights

  • Preserves data privacy by keeping raw samples on-device and sharing only model updates.
  • Exploits large amounts of unlabeled distributed data that would otherwise be wasted.
  • Scales to many heterogeneous clients without requiring label parity across them.
  • Compatible with differential privacy and secure aggregation for additional protection.
  • Reduces annotation cost significantly compared to fully supervised federated learning.

Intuition

This section is available to Pro members. Upgrade to Pro

How it works

This section is available to Pro members. Upgrade to Pro

When to use it

Choose semi-supervised federated learning when data is distributed across devices or institutions that cannot share raw records (privacy, regulation, or bandwidth constraints), and when only a small fraction of that data is annotated. Ideal settings include medical imaging across hospitals, mobile keyboard prediction, and industrial sensor monitoring. Avoid SSFL when you can centralise data without privacy risk — centralized semi-supervised learning is simpler and usually achieves higher accuracy. Also avoid it if all clients have abundant labels (pure federated learning suffices) or if the communication budget is extremely limited (local iterations needed per round are high).

Strengths & limitations

Strengths
  • Preserves data privacy by keeping raw samples on-device and sharing only model updates.
  • Exploits large amounts of unlabeled distributed data that would otherwise be wasted.
  • Scales to many heterogeneous clients without requiring label parity across them.
  • Compatible with differential privacy and secure aggregation for additional protection.
  • Reduces annotation cost significantly compared to fully supervised federated learning.
Limitations
  • Communication overhead is substantial: multiple rounds of upload and download are required.
  • Non-IID (heterogeneous) client distributions can cause global model drift and slow convergence.
  • Pseudo-labels generated locally may be noisy, especially early in training before the model stabilises.
  • Tuning is complex — consistency loss weights, pseudo-label confidence thresholds, and the number of local epochs all interact.

Common pitfalls

This section is available to Pro members. Upgrade to Pro

Applications

This section is available to Pro members. Upgrade to Pro

Frequently asked

How is this different from standard federated learning?

Standard federated learning assumes every client has labeled data and trains purely supervised. SSFL explicitly handles clients with no labels, using semi-supervised objectives (pseudo-labeling, consistency regularization) to extract signal from unlabeled local data.

What happens when client label distributions are very different?

Heterogeneous label distributions (non-IID data) are the main challenge. Techniques such as FedProx, SCAFFOLD, or inter-client consistency losses help stabilize training; without them the global model can drift toward well-labeled clients and perform poorly elsewhere.

Does SSFL provide privacy guarantees?

By design SSFL avoids sending raw data, but model updates can still leak information. Adding differential privacy (clipping and noising gradients) or secure multi-party aggregation provides stronger formal guarantees at some cost to model accuracy.

How many communication rounds are typically needed?

Convergence depends on data heterogeneity, model size, and the number of local epochs. Practical deployments commonly run between 50 and 500 rounds; early stopping based on server-side validation of labeled data is advisable.

Can I use pre-trained models to bootstrap SSFL?

Yes, and it is strongly recommended. Initializing from a strong pre-trained backbone (e.g., a vision or language model) stabilizes early pseudo-labels, accelerates convergence, and substantially reduces the number of rounds required.

Sources

  1. 1.
    Jeong, W., Yoon, J., Yang, E., & Hwang, S. J. (2020). Federated Semi-Supervised Learning with Inter-Client Consistency. International Conference on Learning Representations (ICLR 2021).
    https://arxiv.org/abs/2006.12097
  2. 2.
    Zhang, Z., Chen, Y., Yu, H., & Lu, J. (2021). SemiFed: Semi-supervised Federated Learning with Consistency and Pseudo-Labeling. arXiv preprint arXiv:2108.09412.
    https://arxiv.org/abs/2108.09412

You have read it. What now?

Ask about this methodSimilar methods

Cite this page

ScholarGate. (2026, June 3). Semi-supervised Federated learning. ScholarGate. https://scholargate.app/machine-learning/semi-supervised-federated-learning

Semi-supervised Federated Learning | ScholarGate