Research map / Federated learning and privacy / Personalized and heterogeneous FL

FedCD: Federated Semi-Supervised Learning with Class Awareness Balance via Dual Teachers

AAAI 2024 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Federated medical and multimodal learning · 17 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 FedCD, a federated semi-supervised framework for medical image classification that combines dual (global and local) teacher distillation with knowledge purification and a class awareness balance module to mine rare-class samples from unlabeled clients.

Problem

Federated semi-supervised learning on non-IID medical images suffers from local knowledge shift across clients and class imbalance, and existing methods (e.g., CBAFed, RSCFed) rely on mean teachers or overlook low-confidence rare-class samples.

Approach

Unlabeled clients distill from both a global teacher (receiving the server model) and a local teacher (EMA of client model), with sharpened predictions weighted by KL-based reliability; a class awareness balance module identifies confident classes via a proxy model, flags samples where confident classes rank low as likely rare-class samples, and upweights their loss with a linearly warmed-up facto

Key results

• On HAM10000 (1 labeled, 9 unlabeled clients), FedCD reaches 70.99% accuracy and 83.64% AUC, versus 69.79% / 83.06% for CBAFed.
• On RSNA ICH, FedCD reaches 63.10% accuracy and 79.55% AUC, a 3.76-point accuracy gain over the next best method (CBAFed at 59.34%).
• Ablation on HAM10000: adding CAB alone gives 70.29% accuracy, DTD alone 70.44%, and both together 70.99% (baseline 68.07%).

Measured effects reported in the paper

Surprising findings

• FedIRM and Fed-Consist perform below the lower-bound FedAvg trained on a single labeled client in the non-IID setting.

Limitations

• Both DTD and CAB depend on the warm-up model trained from limited labeled data; scarce labels can impede the server model.
• Some rare-class samples are misidentified, and the method uses hand-set thresholds (tl, th, beta) chosen empirically.

Experimental setup

Datasets: HAM10000 RSNA ICH

Benchmarks: HAM10000 RSNA ICH

Metrics: accuracy AUC precision recall

Baselines: FedAvg (upper bound) FedAvg (lower bound) Fed-Consist FedIRM RSCFed CBAFed

Builds on: FedAvg mean teacher RSCFed CBAFed MixMatch sharpening

Closest papers by meaning

• FedVCK: Non-IID Robust and Communication-Efficient Federated Learning via Valuable Condensed Knowledge for Medical Image Analysis (AAAI 2025 · similarity 0.87)
• MH-pFLID: Model Heterogeneous personalized Federated Learning via Injection and Distillation for Medical Data Analysis (ICML 2024 · similarity 0.85)
• Dual-calibrated Co-training Framework for Personalized Federated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.85)
• Combating Data Imbalances in Federated Semi-supervised Learning with Dual Regulators (AAAI 2024 · similarity 0.85)
• dFLMoE: Decentralized Federated Learning via Mixture of Experts for Medical Data Analysis (CVPR 2025 · similarity 0.85)
• (FL)$^2$: Overcoming Few Labels in Federated Semi-Supervised Learning (NeurIPS 2024 · similarity 0.84)