Research map / Federated learning and privacy / Personalized and heterogeneous FL
Confusion-Resistant Federated Learning via Diffusion-Based Data Harmonization on Non-IID Data
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 CRFed, a federated learning framework for non-IID data that combines an indicator-function-based client/sample weighting, diffusion-based data harmonization, and adaptive learning rates, evaluated on MNIST, FashionMNIST, CIFAR-10/100 and NIPD.
Problem
Federated learning on non-IID data suffers from divergent client updates, model inconsistency, and slow convergence under severe data heterogeneity.
Approach
Uses a loss- and uncertainty-based indicator function to drive client selection, importance-sampling weights, and learning rates, with a diffusion model over the encoded global model plus indicator to harmonize client data distributions.
Key results
- CRFed reaches 0.683 accuracy on CIFAR-10 and 0.389 on CIFAR-100 at beta=0.5, versus 0.650 and 0.362 for FedAvg.
- CRFed achieves 0.882 mAP on NIPD versus 0.821 for FedAvg and 0.871 for the best baseline FedDPMS.
- Ablation shows removing the indicator function drops CIFAR-100 accuracy from 0.389 to 0.365.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CRFed instead of FedDPMS | Federated learning · MNIST | accuracy: 0.983 -> 0.985 | Helped |
| Use CRFed instead of FedDPMS | Federated learning · FashionMNIST | accuracy: 0.876 -> 0.878 | Helped |
| Use CRFed instead of G-FML | Federated learning · CIFAR-10 | accuracy: 0.681 -> 0.683 | Helped |
| Use CRFed instead of FedDPMS | Federated learning · CIFAR-100 | accuracy: 0.386 -> 0.389 | Helped |
| Use CRFed instead of FedDPMS | Federated learning · NIPD | mAP: 0.871 -> 0.882 | Helped |
| Add the indicator function and importance sampling rather than use uniform sampling (ablation) | CRFed · CIFAR-10 | accuracy: 0.661 -> 0.683 | Helped |
| Add the indicator function and importance sampling rather than use uniform sampling (ablation) | CRFed · CIFAR-100 | accuracy: 0.365 -> 0.389 | Helped |
| Add the indicator function and importance sampling rather than use uniform sampling (ablation) | CRFed · NIPD | mAP: 0.861 -> 0.882 | Helped |
| Add diffusion-based data harmonization (ablation) | CRFed · CIFAR-10 | accuracy: 0.670 -> 0.683 | Helped |
| Add diffusion-based data harmonization (ablation) | CRFed · CIFAR-100 | accuracy: 0.373 -> 0.389 | Helped |
| Add diffusion-based data harmonization (ablation) | CRFed · NIPD | mAP: 0.870 -> 0.882 | Helped |
| Use strategic rather than random client selection (ablation) | CRFed · Federated learning | performance | Helped |
Limitations
- Diffusion and adaptive mechanisms may add computational overhead for resource-constrained settings.
- Experiments use only 10 clients with small CNNs and simple benchmark datasets.
- Importance sampling comparison is not fully fair due to differing assumptions among methods.
Experimental setup
Datasets: MNIST FashionMNIST CIFAR-10 CIFAR-100 NIPD
Benchmarks: MNIST FashionMNIST CIFAR-10 CIFAR-100 NIPD
Metrics: test accuracy mAP convergence rounds
Baselines: FedAvg FedProx MOON FedGen HFMDS-FL FRAug G-FML FedCD FedNP FedDPMS FedIR Harmony
Builds on: FedAvg FedProx MOON self-paced learning denoising diffusion probabilistic models
Closest papers by meaning
- Tackling Intertwined Data and Device Heterogeneities in Federated Learning with Unlimited Staleness (AAAI 2025 · similarity 0.89)
- Federated Learning with Sample-level Client Drift Mitigation (AAAI 2025 · similarity 0.88)
- An Aggregation-Free Federated Learning for Tackling Data Heterogeneity (CVPR 2024 · similarity 0.88)
- FedGPS: Statistical Rectification Against Data Heterogeneity in Federated Learning (NeurIPS 2025 · similarity 0.87)
- Fake It Till Make It: Federated Learning with Consensus-Oriented Generation (ICLR 2024 · similarity 0.87)
- FedASMU: Efficient Asynchronous Federated Learning with Dynamic Staleness-Aware Model Update (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.