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

FedVCK: Non-IID Robust and Communication-Efficient Federated Learning via Valuable Condensed Knowledge for Medical Image Analysis

AAAI 2025 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Federated medical and multimodal learning · 13 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 FedVCK, a data-centric federated learning method that condenses each client's knowledge into a small noise-initialized synthetic dataset using latent distribution constraints and model-guided importance sampling, plus relational prototype-wise contrastive learning on the server, to handle severe non-IID medical data under limited communication.

Problem

Federated learning on non-IID medical image data suffers from client drift and poor performance, and existing methods need frequent communication and produce low-quality or repetitive synthesized knowledge.

Approach

Condense local data into a small dataset via distribution matching with fixed per-layer batch-norm statistics as latent constraints, sample training data by model prediction error (importance sampling) so each round captures knowledge the current model lacks, and use hard-negative-aware relational supervised contrastive loss on the server.

Key results

• Under Dir(0.05) with 100 rounds on OrganS (ConvNet), FedVCK reaches 73.23% vs 71.37% for the best prior baseline FedDM; under Dir(0.02) with ResNet18 on Pneumonia it reaches 84.62%.
• Within 10 communication rounds on Path (ConvNet, Dir(0.05)), FedVCK achieves 78.52% vs 72.76% for FedDM (the strongest baseline) and 46.34% for FedAvg at 100 rounds.
• Per-round upload for FedVCK on Path is 2.04 MB vs 12.13 MB for model-based methods such as FedAvg.

Measured effects reported in the paper

Surprising findings

• Model-centric baselines such as FedAvg and FedProx sometimes fall below 20% accuracy on Path under limited rounds, and several data-centric methods (FedGen, DFRD) perform worse than model-centric ones.
• Vanilla distribution matching keeps reducing the condensation loss while the downstream model accuracy does not improve, which motivated the latent distribution constraints.

Limitations

• Evaluated only on 28x28 and 224x224 2D images; extension to 3D CT is left as future work.
• Condensation adds per-client optimization cost (forward/backward passes over T steps) that is not quantified in wall-clock terms.
• Privacy evaluated only with a single membership inference attack (LiRA) and limited AUC gains.

Experimental setup

Datasets: Path OCT OrganS OrganC Pneumonia CIFAR10 STL10 ImageNette

Benchmarks: Path OCT OrganS OrganC Pneumonia CIFAR10 STL10 ImageNette

Metrics: accuracy upload communication cost (MB) membership inference AUC defense rate

Baselines: FedAvg FedProx MOON FedGen FedMix FedGAN DFRD FedDM DESA

Builds on: dataset condensation with distribution matching FedDM SimSiam FedAvg

Closest papers by meaning

• dFLMoE: Decentralized Federated Learning via Mixture of Experts for Medical Data Analysis (CVPR 2025 · similarity 0.89)
• MH-pFLID: Model Heterogeneous personalized Federated Learning via Injection and Distillation for Medical Data Analysis (ICML 2024 · similarity 0.88)
• Think Twice Before Selection: Federated Evidential Active Learning for Medical Image Analysis with Domain Shifts (CVPR 2024 · similarity 0.87)
• FedCD: Federated Semi-Supervised Learning with Class Awareness Balance via Dual Teachers (AAAI 2024 · similarity 0.87)
• Dual-calibrated Co-training Framework for Personalized Federated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.87)
• FedPIA – Permuting and Integrating Adapters Leveraging Wasserstein Barycenters for Finetuning Foundation Models in Multi-Modal Federated Learning (AAAI 2025 · similarity 0.86)