FedRG: Unleashing the Representation Geometry for Federated Learning with Noisy Clients
Authors: Tian Wen, Zhiqin Yang, Yonggang Zhang, Xuefeng Jiang, Hao Peng, Yuwei Wang, Bo Han
Organizations: Institute of Computing Technology, Chinese Academy of Sciences · TMLR Group, Hong Kong Baptist University · The Hong Kong University of Science and Technology · Beihang University
Federated learning (FL) suffers from performance degradation due to the inevitable presence of noisy annotations in distributed scenarios. Existing approaches have advanced in distinguishing noisy samples from the dataset for label correction by leveraging loss values. However, noisy samples recognition relying on scalar loss lacks reliability for FL under heterogeneous scenarios. In this paper, we rethink this paradigm from a representation perspective and propose \method~(Federated under Representation Gemometry), which follows the principle of ``representation geometry priority'' to recognize noisy labels. Firstly, \methodcreates label-agnostic spherical representations by using self-supervision. It then iteratively fits a spherical von Mises-Fisher (vMF) mixture model to this geometry using previously identified clean samples to capture semantic clusters. This geometric evidence is integrated with a semantic-label soft mapping mechanism to derive a distribution divergence between the label-free and annotated label-conditioned feature space, which robustly identifies noisy samples and updates the vMF mixture model with the newly separated clean dataset. Lastly, we employ an additional personalized noise absorption matrix on noisy labels to achieve robust optimization. Extensive experimental results demonstrate that \methodsignificantly outperforms state-of-the-art methods for FL with data heterogeneity under diverse noisy clients scenarios.
Figures & tables
Figure 1: Comparison of noise identification performance between the existing method and FedRG under severe data heterogeneity. The existing methods mainly identify based on the distribution of scalar loss values. In heterogeneous scenarios, the loss value alone becomes unreliable, while FedRG provides more robust spatial geometry to filter out noise samples. Higher FN (False Negatives) but lower TP (True Positives) indicate that noise identification based on loss value wrongly classifies many correct labels as noise labels. The red circle denotes that the noisy samples we want to identify based on the geometric evidence.
Figure 2: Overview of FedRG. The label decoupled spherical representation stage gets the hypersphere of samples. Then, we follow the representation geometry priority principle to split the clean and noisy samples based on the distribution divergence in the vMF distribution. The pseudocode overview of FedRG is provided in Appendix Sec.2
Figure 3: Comparison of the performance among different schemes under four heterogeneous noise scenarios on CIFAR10. Each radar chart corresponds to a specific type of label noise: (a) symmetric with localized noise, (b) symmetric with globalized noise, (c) pairflip with localized noise, and (d) pairflip with globalized noise.
Figure 4: Comparison of model performance under different numbers of selected clients and semantic clusters.
Department of Information Engineering, The Chinese University of Hong Kong · National Key Laboratory on Wireless Communications, University of Electronic Science and Technology of China