FedLabeling: A Federated Framework for Handling Noisy Labeled Data for Efficient Alarm Detection
Satheesh Kumar Perepu, M. Saravanan · 2025
Federated learning (FL) is a collaborative learning method that allows models to be trained without requiring users to transfer their data. However, traditional FL methods assume consistent labeling methods across all users. In real-time applications, especially in the telecommunications sector, different labeling methods across users may lead to label noise in each user’s data, which can impact the overall model performance.To address this issue, existing approaches either assume some users have noise-free labels or require users to report the label noise levels in their data. To circumvent these limitations, we proposed a method called "FedLabeling," which estimates the noise level at each user without needing access to their data. Our method estimates each user’s data impressions from model updates using the zero-shot learning concept and then compares the estimated data with publicly available data to estimate the label noise level.Results on five benchmark datasets from telecom and non-telecom domains demonstrate the improved performance of the global model obtained using our proposed approach compared to traditional methods.