Clustering-based Curriculum Construction for Sample-Balanced Federated Learning

Federated learning is a distributed machine learning scheme that provides data privacy-preserving solution. A key challenge is data distribution heterogeneity of on different parties in federated learning. Existing methods only focus on the training rule of local model rather than data itself. In this paper, we reveal an fact that improving the performance of the local model can bring performance gain to the global model. Motivated by this finding, this paper proposes a Clustering-based curriculum construction method to rank the complexity of instances, and develops a Federation curriculum learning algorithm (FedAC). Specifically, FedAC assigns different weights to training samples of different complexity, which is able to take full advantage of the valuable learning knowledge from a noisy and uneven-quality data. Experiments were conducted on two datasets in terms of performance comparison, ablation studies, and case studies, and the results verified that FedAC can improve the performance of the state-of-the-art Federated learning methods.