Research map / Label-efficient and robust learning / Semi-supervised learning

Mind the Gap: Confidence Discrepancy Can Guide Federated Semi-Supervised Learning Across Pseudo-Mismatch

CVPR 2025 · Label-efficient and robust learning › Semi-supervised learning · approach: Confidence-based pseudo-label selection · 5 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SAGE, a federated semi-supervised learning method that combines local and global model pseudo-labels and softly corrects them using the confidence discrepancy between the two models, which grows with data heterogeneity.

Problem

Federated semi-supervised learning suffers from incorrect hard pseudo-labels that data heterogeneity exacerbates, degrading performance and convergence.

Approach

Use the local model's pseudo-labels when confident and fall back to the global model otherwise (CPG), then soft-correct local pseudo-labels by interpolating with the global model's prediction using an exponential decay of the local-global confidence gap (CDSC).

Key results

• On CIFAR-100 with 10% labels and alpha=1, SAGE reaches 56.06% test accuracy versus 54.18%-ish for the best prior FSSL baseline (FedDB at 51.55 in Table 1 is lower; SAGE improves over FedDB by 3.68 points).
• SAGE reaches 30%/40%/50% target accuracy on CIFAR-100 (alpha=1) in 56/112/242 rounds, about 2.1-2.4x speedup over FixMatch-LPL.
• Adding SAGE as a plugin improves FixMatch on CIFAR-100 by 4.86 to 6.51 points across heterogeneity levels.

Measured effects reported in the paper

Surprising findings

• Hard pseudo-label quality degrades with heterogeneity, and removing incorrect pseudo-labels lets SGD-FSSL approach centralized FixMatch performance.
• Many existing FSSL methods based on hard pseudo-labels perform worse than vanilla FixMatch on some datasets.

Limitations

• Evaluated only in the Label-at-All-Client setting; Label-at-Partial-Client and Label-at-Server not covered.
• Relies on hand-set hyperparameters such as confidence threshold tau=0.95 and sensitivity kappa.

Experimental setup

Datasets: CIFAR-10 CIFAR-100 SVHN CINIC-10

Benchmarks: CIFAR-10 CIFAR-100 SVHN CINIC-10

Metrics: test accuracy convergence rounds pseudo-labeling accuracy

Baselines: FedAvg FedProx FedAvg-SL FixMatch-LPL FixMatch-GPL FedMatch FedLabel FedLoke FedDure FedDB

Builds on: FixMatch FedAvg FlexMatch

Closest papers by meaning

• (FL)$^2$: Overcoming Few Labels in Federated Semi-Supervised Learning (NeurIPS 2024 · similarity 0.92)
• Progressive Distribution Matching for Federated Semi-Supervised Learning (AAAI 2025 · similarity 0.90)
• Combating Data Imbalances in Federated Semi-supervised Learning with Dual Regulators (AAAI 2024 · similarity 0.89)
• BSemiFL: Semi-supervised Federated Learning via a Bayesian Approach (ICML 2025 · similarity 0.89)
• SemiDFL: A Semi-Supervised Paradigm for Decentralized Federated Learning (AAAI 2025 · similarity 0.89)
• FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.88)