Uncertainty-Aware Federated Categorical Dirichlet Allocation for Reliable Industrial Systems

Abstract

For distributed industrial systems, federated learning (FL) can collaborate with a large number of devices for efficient modeling without sharing local data. However, the unreliability of the global model in existing FL methods poses a significant barrier to practical industrial applications. To promote the reliability of FL, we develop an uncertainty-aware FL framework, named federated categorical Dirichlet allocation (FedCDA), that adaptively adjusts to the diverse uncertainty levels inherent in each client’s data. First, we allocate the local client with the Dirichlet distribution to parameterize the categorical prior, thus quantifying the local uncertainty. Second, we calibrate the local Dirichlet distribution by aggregating concentration surrogates (CSs) to eliminate the uncertainty bias resulting from the data heterogeneity. Third, we design a client selection scheme based on the local uncertainty of each client to aggregate reliable clients. Finally, we conduct extensive experiments to demonstrate our method’s superior reliability performance compared to alternative FL algorithms. Our results demonstrate that FedCDA enhances the trustworthiness of the global model across various datasets and confirms its effectiveness in improving both accuracy and reliability.