FBCCNet: a Bayesian perspective of federated learning with crowdsourced annotations on client side

Federated learning has emerged as a promising privacy preserving distributed learning framework, enabling participating clients to collaboratively train a shared global model without exposing their local data. Existing federated learning algorithms assume that each local device possesses high-quality labels provided by an experienced expert. However, in many real-world applications, data labels on clients are annotated by multiple crowd members with varying levels of expertise, leading to crowdsourced data with partial labeling, probabilistic worker participation, and variable reliability levels. To address these challenges, in this paper, we propose a novel federated Bayesian classifier combination neural network named FBCCNet that simultaneously aggregates crowdsourced labels and collaboratively trains a global model in a distributed manner. By integrating the local variational EM algorithm with a stopping criterion computed from the aggregated evidence lower bound, the distributed learning performance on crowdsourced data is significantly enhanced. Empirical results demonstrate that our proposed FBCCNet outperforms the conventional majority voting and expectation-maximization approaches and is robust to varying crowdsourced labeling ratios on both independent and identically distributed (IID) and non independent and identically distributed (non-IID) data.