FNRE: A Novel Approach to Heterogeneous Label Noise Rates Estimation in Federated Learning

Federated learning (FL) enables collaborative model training across decentralized clients without sharing raw data, thereby preserving privacy. However, in real-world FL deployments—such as sensor-based activity recognition, wearable health monitoring, and industrial Internet of Things, where local training data often suffer from heterogeneous noisy labels due to diverse collection environments, sensor limitations, and labeling errors. These noisy labels, typically distributed unevenly across clients due to differences in client-side annotation, exacerbate Non-Independent and Identically Distributed (non-IID) data issues, leading to biased updates, unstable convergence, and degraded global model performance. Accurate estimation of client-specific noise rates is therefore crucial for adaptive algorithm selection, personalized parameter tuning, noise-aware aggregation, and resource allocation in FL. Existing noise rate estimation methods, primarily developed for centralized settings, require client-specific clean validation sets or prior knowledge of noise, making them impractical in privacy-sensitive federated settings. In this work, we propose a federated noise rate estimation (FNRE) method that eliminates the need for per-client clean datasets or prior knowledge of noise. Our approach requires only a minimal assumption—at least one client with a small clean validation set—and leverages the global model’s predictions to estimate local noise rates across all clients. Specifically, the method computes global prediction accuracy using data from the small, clean subset of clients, broadcasts this accuracy to all participants, and enables each client to infer its noise rate using its own annotated labels and the predicted label sequence. We further provide a theoretical analysis with provable error bounds. Extensive experiments on image classification (CIFAR-10, CIFAR-100) and sensor-based activity recognition (Widar, WISDM-W) under various synthetic and real-world noisy label settings demonstrate that our method achieves a noise rate estimation Mean Absolute Error (MAE) of only 0.82%–2.19%, outperforming state-of-the-art baselines by 29.8%–49.9% on average while maintaining practicality in privacy-sensitive federated environments.