FedCova: Robust Federated Covariance Learning Against Noisy Labels

Xiangyu Zhong, Xiaojun Yuan, Ying-Jun Angela Zhang

Keywords: federated learning, noisy labels, feature learning, robust learning model, mutual information maximization

TL;DR: We propose FedCova, a novel federated learning framework that enhances robustness to noisy labels by aligning and leveraging the covariance structure of features across distributed devices.

Abstract: This paper addresses the critical challenge of federated learning (FL) under noisy labels by exploiting intrinsic robustness grounded in covariance structures. Noisy labels in distributed datasets induce severe local overfitting and consequently compromise the global model in FL. Most existing solutions rely on selecting clean devices or aligning with public clean datasets, rather than endowing the model itself with robustness. In this paper, we propose *FedCova*, a noise-resistant federated covariance learning framework, to enhance the model's intrinsic robustness via a new perspective on feature covariances. Specifically, FedCova encodes data into a discriminative but resilient feature space to tolerate label noise. Built upon mutual information maximization, we design a novel objective for federated lossy feature encoding, which is driven solely by the feature covariances of different classes with an error tolerance term. Leveraging feature subspaces characterized by covariances, we construct a subspace-augmented federated classifier. FedCova unifies three key processes through the covariance: (1) training the network for feature encoding, (2) constructing a classifier directly from the learned features, and (3) correcting labels based on feature subspaces. The server aligns the federated classifier via covariance aggregation, which devices use to build local external correctors for relabeling, avoiding self-correction. We implement FedCova under heterogeneous data distribution across various noisy settings. Experimental results on CIFAR-10/100 and real-world noisy dataset Clothing1M demonstrate the superior robustness of FedCova compared with the state-of-the-art methods.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 12237