FedRG: Unleashing the Representation Geometry for Federated Learning with Noisy Clients

[CVPR 2026]

This repository contains the official implementation of "FedRG: Unleashing the Representation Geometry for Federated Learning with Noisy Clients".

Overview

Federated learning under noisy supervision is hard for a simple reason: in heterogeneous client environments, scalar loss is no longer a reliable signal for noisy-sample detection. FedRG rethinks noisy-label learning from the representation side. It first learns label-decoupled spherical representations with self-supervised training, then fits a vMF-based geometry model on the hypersphere, compares label-free geometric evidence with label-conditioned evidence, and finally performs robust local optimization with a personalized noise absorption matrix.

Highlights

• A two-stage federated pipeline for noisy-client learning:
  • Stage I: SimCLR-based federated pretraining for label-decoupled spherical representations.
  • Stage II: geometry-based noisy-sample identification and robust optimization.
• A representation geometry priority principle that avoids relying directly on unstable small-loss heuristics.
• A client-specific noise absorption matrix for robust training on noisy subsets.

Main Experimental Setting

We evaluate FedRG on CIFAR-10, CIFAR-100, and SVHN. Unless otherwise stated, the data are partitioned with a Dirichlet distribution with concentration parameter alpha = 0.1, and the local label noise rate is fixed to epsilon = 0.4. We consider both globalized and localized corruption, each under symmetric and pairflip noise. Most experiments use K = 10 clients, and we additionally report results on CIFAR-10 with K = 100 clients.

Default training settings used in the paper:

• Optimizer: SGD
• Learning rate: 0.01
• Momentum: 0.9
• Weight decay: 5e-4
• Batch size: 64
• Stage I rounds: 150
• Stage II rounds: 350
• Number of semantic clusters |G|:
  • CIFAR-10: 10
  • CIFAR-100: 50
  • SVHN: 10

Repository Structure

The public codebase is organized around the following components:

.
├── conf/                    # experiment configs
├── data_preprocessing/      # dataset construction, partitioning, and noise injection
├── model/                      # backbone networks and model manager
├── loss/                       # loss functions, including SCE
├── utils/                      # metrics, logging, recorders, helper functions
├── baseFedAvg/                 # federated training framework
├── FedRG/                      # FedRG-specific client/server/manager logic
├── main.py                   # runnable experiment scripts
└── README.md

If your local repository uses different names, please replace this section with the exact released layout before making the repository public.

Citation

If you find this repository useful, please cite:

@inproceedings{wen2026fedrg,
  title={FedRG: Unleashing the Representation Geometry for Federated Learning with Noisy Clients},
  author={Tian Wen and Zhiqin Yang and Yonggang Zhang and Xuefeng Jiang and Hao Peng and Yuwei Wang and Bo Han},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2026}
}

Contact

For questions about the paper or code, please contact us marrowd611@gmail.com.